computer science & information technology 41 · computer science & information technology...

Computer Science & Information Technology 41

David C. Wyld Natarajan Meghanathan (Eds)

Computer Science & Information Technology

The Seventh International Conference on Wireless & Mobile Network (WiMo 2015) Sydney, Australia, May 23 ~ 24 - 2015

AIRCC

Volume Editors

David C. Wyld, Southeastern Louisiana University, USA E-mail: [email protected] Natarajan Meghanathan, Jackson State University, USA E-mail: [email protected] ISSN: 2231 - 5403 ISBN: 978-1-921987-38-0

DOI : 10.5121/csit.2015.51001 - 10.5121/csit.2015.51008

This work is subject to copyright. All rights are reserved, whether whole or part of the material is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the International Copyright Law and permission for use must always be obtained from Academy & Industry Research Collaboration Center. Violations are liable to prosecution under the International Copyright Law. Typesetting: Camera-ready by author, data conversion by NnN Net Solutions Private Ltd., Chennai, India

Preface

The Seventh International Conference on Wireless & Mobile Network (WiMo-2015) was held in Sydney, Australia, during May 23~24, 2015. The Fourth International Conference on Information Technology Convergence and Services (ITCSE-2015), The Fourth International conference on Advanced Computer Science and Information Technology (ICAIT-2015), and The Third International Conference of Networks and Communications (NC-2015) were collocated with the WiMo-2015. The conferences attracted many local and international delegates, presenting a balanced mixture of intellect from the East and from the West. The goal of this conference series is to bring together researchers and practitioners from academia and industry to focus on understanding computer science and information technology and to establish new collaborations in these areas. Authors are invited to contribute to the conference by submitting articles that illustrate research results, projects, survey work and industrial experiences describing significant advances in all areas of computer science and information technology. The WiMo-2015, ITCSE-2015, ICAIT-2015, NC-2015 Committees rigorously invited submissions for many months from researchers, scientists, engineers, students and practitioners related to the relevant themes and tracks of the workshop. This effort guaranteed submissions from an unparalleled number of internationally recognized top-level researchers. All the submissions underwent a strenuous peer review process which comprised expert reviewers. These reviewers were selected from a talented pool of Technical Committee members and external reviewers on the basis of their expertise. The papers were then reviewed based on their contributions, technical content, originality and clarity. The entire process, which includes the submission, review and acceptance processes, was done electronically. All these efforts undertaken by the Organizing and Technical Committees led to an exciting, rich and a high quality technical conference program, which featured high-impact presentations for all attendees to enjoy, appreciate and expand their expertise in the latest developments in computer network and communications research.

In closing, WiMo-2015, ITCSE-2015, ICAIT-2015, NC-2015 brought together researchers, scientists, engineers, students and practitioners to exchange and share their experiences, new ideas and research results in all aspects of the main workshop themes and tracks, and to discuss the practical challenges encountered and the solutions adopted. The book is organized as a collection of papers from the WiMo-2015, ITCSE-2015, ICAIT-2015, NC-2015

We would like to thank the General and Program Chairs, organization staff, the members of the Technical Program Committees and external reviewers for their excellent and tireless work. We sincerely wish that all attendees benefited scientifically from the conference and wish them every success in their research. It is the humble wish of the conference organizers that the professional dialogue among the researchers, scientists, engineers, students and educators continues beyond the event and that the friendships and collaborations forged will linger and prosper for many years to come.

David C. Wyld

Natarajan Meghanathan

Organization

General Chair

Natarajan Meghanathan Jackson State University, USA Dhinaharan Nagamalai Wireilla Net Solutions PTY LTD, Australia

Program Committee Members

Abdelhafid Abdelmalek Tlemcen University, Algeria Ahmed Mohamed Khedr University of Sharjah, UAE Ali Abid D. Al-Zuky Mustansiriyah University, Iraq Ali Chaabani National School of Engineering, Tunisia Ali Hussein Alexandria University, Egypt Ankit Chaudhary Truman State University, USA Barbaros Preveze Cankaya University Ankara, Turkey Bela Genge Petru Maior University of Tg. Mures, Romania Chun-Yi Tsai National Taitung University, Taiwan Dac-Nhuong Le Haiphong University, Vietnam David C. Wyld Southeastern Louisiana University, USA Emilio UR University of La Rioja, Spain Farhan University Of Indonesia, Indonesia Farzad Kiani Istanbul Sabahattin Zaim University, Turkey Foudil Cherif Biskra University, Algeria Girija Chetty University of Canberra, Australia Grienggrai Rajchakit Maejo University, Thailand Gullanar M Hadi Salahaddin University, Hawler, Iraq Habib Rasi Shiraz University of Technology, Iran Hamadouche M Hamadouche Maamar, Algeria Hamdi Hassen Sfax University, Tunisia Hamdi M National Engineering School of Tunis, Tunisia Héldon José Professor of Integrated Faculties of Patos, Brazil Ian Tan Multimedia University, Malaysia Iram Siraj Aligarh Muslim University, India Isa Maleki Islamic Azad University, Iran Israashaker Alani Minustry If Science and Technolgy, Iraq Jacques Epounde Ngalle Robert Morris University, USA Jose Enrique Armendariz-Inigo Public University of Navarre, Spain Koushik Majumder West Bengal University of Technology, India Laudson Souza Professor of Integrated Faculties of Patos, Brazil Li, Zheng University of Bridgeport, USA Majlinda Fetaji South East European University, Macedonia Mehdi Nasri Islamic Azad University, Iran Meyyappan T Alagappa University, India Moez Hizem Sup'Com, Tunisia Mohamed Elboukhari ESTO, Oujda, Morocco

Mohamed Hassan American University of Sharjah, UAE Mohamed Khamiss Suez Canal University, Egypt Mohammed Amin Higher Colleges of Technology, UAE Muhammad Ali University of Bradford, United Kingdom Muhammad Sajjadur Rahim University of Rajshahi, Bangladesh Neeraj Kumar Nalanda College of Engineering, India Paramartha Dutta Visvabharati University, India Pinaki Bhaskar The National Research Council, Italy Polgar Zsolt Alfred Technical University of Cluj Napoca, Romania Raed I Hamed University of Anbar Ramadi, Iraq Reza Ebrahimi Atani University of Guilan, Iran Roheet Bhatnagar Manipal University, India Saad Darwish University of Alexandria, Egypt Saadat Pourmozafari Tehran Poly Technique, Iran Salem Nasri Qassim University, Kingdom of Saudi Arabia Santhi Balaji Bangalore University, India Sattar B.Sadkhan University of Babylo, Iraq Seyed Hossein Hosseini Nazhad Islamic Azad University, Iran Seyyed Reza Khaze Islamic Azad University, Iran Sharma Chakravarthy University of Texas at Arlington, USA Shuxiang Xu University of Tasmania, Australia Simi Bajaj University of Western Sydney, Australia Sokyna Qatawneh Al-Zaytoonah University Of Jordan, Jordan Taruna S Banasthali University, India Tomer Turkiye Istanbul S.Zaim University, Istanbul, Turkey Utku KÃ–SE Usak University, Turkey Wajeb Gharibi Jazan University, Saudi Arabia Zuhal Tanrikulu Bogazici University, Turkey

Technically Sponsored by

Networks & Communications Community (NCC)

Computer Science & Information Technology Community (CSITC)

Digital Signal & Image Processing Community (DSIPC)

Organized By

Academy & Industry Research Collaboration Center (AIRCC)

TABLE OF CONTENTS

The Seventh International Conference on Wireless & Mobile Network

(WiMo 2015)

Performance Evaluation of Routing Protocols for Delay Tolerant

Networks …………………………………………………………………………... 01 - 12

Luming Wan, Feiyang Liu, Juan Zhang and Haibo Zhang

Modeling Social Gauss-Markov Mobility for Opportunistic Network ..........… 13 - 25

GuoDong Kang, Jinzhi Ning and GuoLiang Kang

The Fourth International Conference on Information Technology

Convergence and Services (ITCSE 2015)

A Triangle-Triangle Intersection Algorithm …………………………….……... 27 - 35

Chaman L. Sabharwal and Jennifer L. Leopold

Fourth International conference on Advanced Computer Science and

Information Technology (ICAIT 2015)

Design and Implementation of Intel-Sponsored Real-Time Multiview

Face Detection System ……………………………………………….…….……... 37 - 47

Bohua Gan, Vincent Chang, Guanying Wang, Xiuli Pan, Guan Wang,

Naihai Zou and Felming Feng

Enhancing Performance of an HPC Cluster By Adopting Non-Dedicated

Nodes …………………………………………………….………………….……... 49 - 59

Pil Seong Park

Adaptive Authentication : A Case Study for Unified Authentication

Platform ……………………………………………...……….…………….……... 61 - 72

Khairul Azmi Abu Bakar, Nor Izyani Daud and Mohd Shafeq Md Hasan

Third International Conference of Networks and Communications

(NC 2015) Sensor Selection Scheme in Wireless Sensor Networks : A New Routing

Approach …………………………………………………….……….…….……... 73 - 79

Mohammad Alwadi and Girija Chetty

An Group Behavior Mobility Model for Opportunistic Networks ….….……... 81 - 89

GuoDong KANG and GuoLiang KANG

David C. Wyld et al. (Eds) : WiMo, ITCSE, ICAIT, NC - 2015 pp. 01–12, 2015. © CS & IT-CSCP 2015 DOI : 10.5121/csit.2015.51001

PERFORMANCE EVALUATION OF

ROUTING PROTOCOLS FOR DELAY

TOLERANT NETWORKS

Luming Wan1, Feiyang Liu1, Juan Zhang2, and Haibo Zhang1

1Department of Computer Science, University of Otago, Dunedin, New Zealand {lwan, feiyang, haibo}@cs.otago.ac.nz

2School of Reliability and System Engineering, Beihang University, Beijing, China

[email protected]

ABSTRACT.

Delay Tolerant Network (DTN) is a promising technology which aims to provide efficient

communication between devices in a network with no guaranteed continuous connectivity. Most

of the existing routing schemes for DTNs achieve message delivery through message replication

and forwarding. However, due to the lack of contemporaneous end-to-end communication path,

designing routing protocols that can achieve high delivery rate with low communication

overhead is a challenging problem. Some routing protocols appear with high similarity, but

their performance are significantly different. In this paper, we evaluate several popular routing

protocols in DTNs, including Epidemic, Spray and Wait, PRoPHET, and 3R through extensive

trace-driven simulations. The objective is to evaluate the performance of different routing

schemes using different data traces and investigate the optimal configuration setting for each

routing scheme. This paper provides important guidances on the design and selection of routing

protocols for given delay tolerant networks.

KEYWORDS

Delay Tolerant Network, Performance Evaluation, Routing

1. INTRODUCTION

The increasing popularity of ubiquitous computing and communication leads to a huge desire on data exchanging between wireless mobile devices, e.g. cell phones, laptops, tablets, and other carriable devices, regardless of whether any guaranteed end-to-end connection exists. Delay tolerant networks (DTNs) address the technical issue on communication between devices that lose continuous connectivity due to mobility. Currently, DTNs have been applied to a vast of areas, including vehicular networks[1], wildlife tracking[2], and social network analysis[3], etc. However, DTNs are still appearing with numerous of limitations. For example, due to the lack of synchronous end-to-end connectivity in DTNs, mobile devices have to carry the messages and forward them opportunistically upon encountering the destinations, or forward messages to other relays to help the delivery. This could result in incredible long transmission delay and high error

2 Computer Science & Information Technology (CS & IT)

rate. In addition, the constraints of wireless devices, such as storage capacity, communication bandwidth and battery power, can significantly impact the delivery rate. Hence, routing protocols for DTNs should be able to adapt to the network variation and efficient to make use of the hardware resources. Recent studies exhibit that most of the existing routing protocols designed for DTNs appear with high similarity in concepts, but their performance are significantly disparate. Epidemic[4] performs robust and blind flooding of messages to the network. Spray and Wait[5] is a flooding-controlled version of Epidemic, as it simply limits the amount of messages that can be flooded to the network. In most prediction-based schemes, such as PRoPHET[6], MaxProp[7] and PER[8], the message forwarding decisions are made based on a quality metric called encounter predictability which is accumulated from each time of encounter. However, the above prediction-based schemes ignore some important encounter information, such as the contact time. Some studies demonstrate that human activities appear with high repetition, such as weekly meetings. The performance of routing schemes could be enhanced significantly through the exploitation and utilization of those regular patterns. 3R[9] is a fine-grained history-based routing scheme, by which the contact time of encounters are perfectly recorded. Due to the maintenance of fine-grained encounter history, the encounter predictability in 3R is time-dependent, that is, it can be calculated based on only the past contacts that occurred in the same period as the lifetime of the packet, rather than making a long-term average estimation as in most prediction-based protocols, such as PRoPHET. In addition, 3R is a forwarding-based scheme, each node always forwards the original message to the next good relay rather than transmitting replicas as in all the protocols above. However, single-copy for each message is not ideal for DTNs since message could be easily lost due to unstable connectivity and other uncertainties. It is difficult to design an absolutely perfect routing protocol for all DTN applications. The performance of a routing protocol can be affected by a large number of factors, such as the popularity and active rate of nodes, different setting of parameters for mobile devices, etc. In this paper, we evaluate the performance of several most popular routing protocols, including Epidemic, Spray and Wait, PRoPHET, and 3R through trace-driven simulations. Also, we investigate the impact of different parameters of Spray and Wait and PRoPHET to achieve their optimal settings. The contribution of this paper is to provide some important guidances on routing protocol design and selection for delay tolerant networks. The reminder of this paper is organized as follows: Section 2 introduces the protocols we are going to evaluate in detail. Section 3 introduces the data traces and experiment setup. Section 4 shows the performance evaluation in two perspectives. Firstly, we present the impact of different parameter settings for Spray and Wait, and PRoPHET protocols. Then, we evaluate the performance of all the protocols with their optimal parameter settings. The paper is concluded in Section 5.

2. ROUTING PROTOCOLS

This section introduces four protocols of our comparisons in detail, which are Epidemic, PRoPHET, Spray and Wait, and 3R respectively.

Computer Science & Information Technology (CS & IT) 3

2.1. Epidemic

Epidemic[4] is a pure flooding-based routing protocol for DTNs. Each device (source or relaying node) always propagates message replicas to all of the contactable nodes until the message is received by the destination node or the message deadline expires. Due to the nature of unlimited flooding, Epidemic is able to achieve the best message delivery rate when each node has an infinite memory buffer, but its performance deteriorates significantly when each node only has limited resources. This is because numerous of message replicas have to be dumped due to the memory overflow.

2.2. Spray and Wait

Similar to Epidemic, Spray and Wait [5] is also a flooding-based routing protocol, but it controls the message flooding by directly limiting the number of replicas that an original message can produce. Each message can only propagate a constant amount of replicas in the spray phase. After that, the original message and all of its replicas switch to the wait phase, that is, those messages only wait for the destination node encountering their holders rather than further spreading replicas to other relays. Binary Spray and Wait is a derivation of original version, as nodes spread the replicas of a message in a binary tree manner. Suppose a message is allowed to spread L replicas, the source node will generate L replicas in the beginning and spread them among encountered relays. Suppose any node A that has n(1 < n< L) replicas of message m, and

it encounters a node B with no replica of $m$ previously. A will send 2n replicas to B and keep

the rest of 2n replicas to wait for more relaying nodes. When only one replica is left in A or any

other relaying node, they stop spreading any replicas and wait for encountering the destination node. Spray and Wait can achieve a trade-off between the delivery rate and transmission overhead by setting a proper limitation of L.

2.3. PRoPHET

PRoPHET[10] is a prediction-based scheme and it is one of the few DTN routing protocols that have an IETF draft. It implements a quality metric called encounter predictability to measure the capability of the encountering nodes whether or not it can transmit the message to the destination. A replica of a message is propagated to the encountering node if it has a higher predictability than its holder. This guarantees that each time a message can always be propagated to a better relay. The encounter predictability is integrated in three perspectives, which are direct probability, transitivity and ageing. Direct probability is the probability of transmitting a message directly between two nodes. It is updated whenever two nodes directly encounter each other as follows:

��, �� = ��, �� + 1 − � − ��, �� × �� , (1)

where Pold(a,b) is the encounter probability of node a and b before the current encounter occurs,

Pencounter∈[0,1]is a scaling factor at which the probability increases on encounters, and δis a small

positive value to set an upper bound on P(a,b). Transitivity estimates the probability of indirect contact that through multiple-hop relays, as shown below:


��, �� = ��, �� + 1 − ��, �� × ��, �� × ��, �� × �, (2)

Where P(a,c)is the encounter probability between node a and c through relay node b, Pold(a,c) is

the probability from previous time of update, and β∈ [0,1] is a scaling factor that decides how

large the impact of transitivity should have on the encounter predictability. PRoPHET uses an ageing mechanism to decay the encounter predictability and eliminate the long-time inactive nodes, as shown below:

��, �� = ��, �� × ��, (3)

where �∈ [0,1]is the ageing constant, and K is the number of time units elapsed since the last

timeP(a,b) was aged. As shown in the above Equations, there are three important parameters for PRoPHET protocol: scaling factors for direct delivery Pencounter and transitivity β, and the ageing constant �. They have significant influence on the performance and overheads of Prophet scheme. PRoPHET can be an efficient routing protocol by setting the proper parameters if the encountering patterns in DTNs are predictable, because the messages and replicas are only forwarded to the relays with more benefits on message delivering. At present, there are two versions of PRoPHET protocol: PRoPHET'03[6] and PRoPHET'12[10]. They have different recommendation settings for these parameters, while the principles of message forwarding for both versions are basically identical.

2.4. 3R 3R[9] characterizes the encounter history in a fine-grained form by storing more information for each encounter, including the nodes that the encounter occurs among, the start time and the end time of this encounter. Each node maintains a table for the encounter information of its contacted nodes. The fine-grained history in a node will group the contacts according to different types of the day {weekday, weekend}. Each item in the table represents a fixed time interval of the real world (e.g. 1 hour), and it records the overall contact frequency between the table holder and the specified node at exact the same period in the past. Suppose a message m has a lifetime that spansk slots. Let F

a

i be the number of contacts of node a that occurred at a time slot i in the past, and F

i

ab be the number of contacts occurred between node a and b at time slot i, the estimated contact probability p

i

ab at slot i is:

�� = !"#

$

!"$ . (4)

Suppose the lifetime of a message m spans k time slots, the overall probability that message m can be delivered from node a to node b before it expired is estimated by integrating the probabilities of these k slots as follows:

��%� = 1 − ∏ �1 − ��

'( . (5) Due to the maintenance of fine-grained history, a time-dependent forwarding prediction is enabled by only figuring out the delivery probability within the lifetime of a message, rather than calculating a long-term average estimation. Similar to PRoPHET, 3R also compares the


predictability and forwards messages to nodes with higher delivery probability. Whereas, 3R only directly forwards the original message to the next relay, instead of sending a replica as in PRoPHET. Due to the predictability calculation based on fine-grained contact history and single-copy message forwarding, 3R can effectively reduce delivery overheads with no much loss on message delivery rate.

3. DATA TRACES AND EXPERIMENT SETUP

3.1. Data Traces

We evaluate above four representative routing protocols for DTNs with extensive trace-driven simulations. To increase the accuracy and reliability of our evaluation, we use two realistic DTN data traces: INFOCOM[11] and MIT Reality[12], which are obtained from the open-source website, CRAWDAD[13].

• INFOCOM: This data trace consists of Bluetooth sightings and it records 4 days of contacts occurred during the INFOCOM'05 conference. In the experiment, 24 internal devices monitored the network and more than 200 external devices were discovered. Each device had a scan granularity of 120 seconds, and each scan lasted for 5 seconds.

• MIT Reality: This trace contains the contact information including communication, proximity, and location from faculties and students at MIT over the course of the 2004-2005 academic year. To collect this trace, 89 devices recorded the contacts, and more than 20,000 devices were discovered. Each device scanned to discover neighbors every 5 minutes. In our simulations, we use the devicespansubtrace that records Bluetooth contacts for 1 month.

3.2. Experiment Setup

In our simulations, the routing schemes are evaluated using the same message trace and contact trace. To avoid the negligible delivery rate caused by numerous long-time inactive nodes which actually seldom communicate, we randomly select the source node and the destination node of a message from 20 most active nodes during the lifetime of this message. In the INFOCOM trace, each source generates a message with the probability of 0.15 in every 600 seconds, whereas the probability is 0.6 in the MIT Reality trace. The lifetime of each message varies from 2 hours to 3 days if not specially notified. The size of each generated message is randomized from 2k bytes to 100k bytes. For the MIT Reality trace, the network is warmed up for one week at the beginning of each simulation run. This avoids the inaccurate prediction on encounter probability for prediction-based protocols, i.e.PRoPHET and 3R, when the network is just booted up. Also, we reserve 3 days at the end of simulations to avoid any message left in the network after the simulation is completed. During these warm-up and sinking periods, messages are not allowed to be generated. This process is not available for the INFOCOM data trace, since it only contains 4 days of contact traces.


4. PERFORMANCE EVALUATION

In this section, we firstly exhibit the performance of Spray and Wait and PRoPHET with different parameter settings using the INFOCOM data trace, since they have some specific parameters. Then we investigate the performance of all four protocols, i.e. Epidemic, Spray and Wait, PRoPHET, and 3R, through trace-driven simulations. The performances of these four protocols are investigated in two perspectives: different memory buffer sizes and packet lifetimes. The performance is evaluated using the following three metrics:

• average message delivery rate: the proportion between the number of successfully delivered messages against the number of original messages.

• message overhead (MO): the ratio of the number of replicas against the number of original messages.

• communication overhead (CO): the ratio of the difference between the number of delivered messages and hop-by-hop transmissions to the number of delivered messages.

The parameter setting for Spray and Wait is slightly different in each data trace in accordance with the characteristic of the data trace. The Binary Spray and Wait protocol is implemented in the entire experiments as it performs slightly better than its original version. Due to the huge difference on popularity of nodes in these two traces, the maximum number of replicas for Binary Spray and Wait with INFOCOM trace is set to 5, and 70 with MIT Reality trace. The parameters in PRoPHET are configured with the values provided in PRoPHET'12[10] by default. All the nodes are set with infinite memory buffer if not specified. All the protocols dump messages using the First In First Out policy when the message buffer is full, that is, the first message in the buffer queue (oldest) is dumped to prevent buffer overflow.

4.1. Impact of Parameters for Spray-and-Wait and PRoPHET

Spray-and-Wait: Fig. 1 shows the performance of the original and binary version of Spray and Wait routing protocol with different maximum number of replicas. In Fig.1(a), it can be seen that the original Spray and Wait achieves higher delivery rate than the binary version when the maximum number of replicas is less than 10. With the further increase on the number of replicas, the binary version performs better than the original version. That is because the original scheme can only spread the replicas around the neighbours, while the binary scheme can spread the replicas more widely in a binary-tree manner if enough number of replicas are allowed. We can seethe delivery rate of both protocols remain steady when the maximum replicas exceeds 40. The highest average delivery rate achieved by the original and the binary versions are 0.9314 and 0.9382, respectively. That means the number of replicas are enough for two Spray and Wait protocols with no further benefits for more replicas. Meanwhile, MO and CO for both schemes appear with the similar increasing trend as shown in Fig. 1(b) and (c). The MO and CO of the original Spray and Wait increase apparently with the maximum replicas less than 40, and they stay constant afterwards also due to the limitation on the number of neighbours, whereas both two kinds of overheads of Binary Spray and Wait grow almost linearly since the replicas can be spread to more relay nodes on the network. From these figures, it can be seen that with a large number of replicas, the multi-hop routing of Binary Spray and Wait is able to perform better than the only two-hop routing in original version, but it suffers almost twice of overheads than the original version.


Fig. 1. The performance of Spray and Wait with different number of message replicas, (a) delivery rate, (b) message overhead, (c) communication overhead.

The highest average delivery rate achieved by the original and the binary versions are 0.9314 and 0.9382, respectively. That means the number of replicas are enough for two Spray and Wait protocols with no further benefits for more replicas. Meanwhile, MO and CO for both schemes appear with the similar increasing trend as shown in Fig. 1(b) and (c). The MO and CO of the original Spray and Wait increase apparently with the maximum replicas less than 40, and they stay constant afterwards also due to the limitation on the number of neighbours, whereas both two kinds of overheads of Binary Spray and Wait grow almost linearly since the replicas can be spread to more relay nodes on the network. From these figures, it can be seen that with a large number of replicas, the multi-hop routing of Binary Spray and Wait is able to perform better than the only two-hop routing in original version, but it suffers almost twice of overheads than the original version. PRoPHET: The performance of PRoPHET routing protocol is critically impacted by the parameters shown in Table 1. PRoPHET provides totally different recommended parameters in their 2012 version [10] compared to the 2003 version [6].

Table 1.PROPHET recommended parameters

Parameters Pencounter β γ

PRoPHET 2003 0.75 0.25 0.98

PRoPHET 2010 0.5 0.9 0.999

Fig. 2 shows the performance of PRoPHET'03 and PROPHET'12. In these experiments, each node has infinite memory buffer to store messages and replicas. The principles of message forwarding for these two versions are basically identical. It can be seen that in Fig. 2(a), the delivery rate is increased by 0.5% with the parameters of PRoPHET'12 in comparison with PRoPHET'03. A larger ageing parameterγmakes PROPHET'12 better to tolerate the variation of contact patterns and achieve higher delivery rate. However, both MO and CO are raised by 6.1% and 5.4% respectively as shown in Fig. 2(b). This is because a largerβin PRoPHET'12 leads to a higher impact on the transitive connectivity, and a lower Pencounter reduces the impact of direct delivery. Hence, more replicas need to be generated and forwarded by the relay nodes.


4.2. Performance Evaluation with Different Memory Sizes

Fig.2.PRoPHET'03 versus PRoPHET'12, (a) delivery rate, (b) overheads.

In each mobile device, the memory buffer used for message delivery always has a limited size. The available memory has huge influence on the delivery rate and overheads. We evaluate the performance of four routing protocols with different memory sizes using INFOCOM and MIT Reality traces. (1) INFOCOM Trace: In Fig. 3(a), we evaluate the influence of memory size on the delivery rate for different routings. It can be seen that the delivery rate of all four routing protocols have a sharp increasing trend with a small memory buffer (<10MB), and then they tend to remain stable even if the memory buffer size keeps growing. Binary Spray and Wait has more outstanding performance than other three routings when each node has a memory buffer of 80M bytes or less. It reaches its maximum delivery rate of 0.923 with the memory buffer size increasing to 60M bytes, and the delivery rate tends to be stable afterwards. That is because it is a controlled flooding scheme (5 replicas) and has less message dropped due to memory overflow than other routings. When the memory size exceeds 60MB, there will be no message dumps for this routing and the delivery rate will be stable. Similarly, 3R also has better performance with a smaller memory buffer size compared with Epidemic and PRoPHET. It has less requirement on memory size because of single-message and its limitation on message forwarding. However, it reaches the maximum delivery rate of 0.792 at the memory size of 40M bytes, which is much smaller compared with the other three protocols. That is because 3R only forwards the original message. Although the delivery rates of Epidemic and PRoPHET raise slowly, their maximum delivery rates are higher than 3R and Spray and Wait (0.941 and 0.930 respectively) since they are memory-hungry schemes. Epidemic protocol needs large amounts of memory for unlimited flooding and can achieve the highest delivery rate with enough memory. PRoPHET also needs a large memory size for transmitting the amounts of messages and replicas. Fig. 3(b) and (c) evaluate the overheads with different memory sizes. We can see that Epidemic always has the highest MO and CO due to the nature of unlimited flooding. Since it needs to generate more replicas when encountering relay nodes with high predictability, PRoPHET also has higher overheads, which is almost half of the overheads as Epidemic. Because the maximum number of replicas in Binary Spray and Wait is set to 5, the maximum MO and CO of Binary Spray and Wait scheme tend to 5 even with a much larger memory size. Since 3R is a single-message forwarding-only scheme, there is no message overhead. The CO of 3R is also lowest because the message is only forwarded to a node with higher predictability according to fine-


grained encounter history. The CO of Binary Spray and Wait and 3R appear with a decreasing trend as the memory size increases, since they have limitations on message forwarding and a larger memory size increases the number of delivered messages by reducing message dumps.

Fig.3.Performance evaluation with different memory sizes using INFOCOM trace,(a) delivery rate, (b) message overhead, (c) communication overhead.

Fig. 4.Performance evaluation with different memory sizes using MIT Reality trace, (a) delivery rate, (b) message overhead, (c) communication overhead.

(2)MIT Reality Trace: As shown in Fig. 4(a), the delivery rate of each routing protocol has similar trend as the memory size increases. However, PRoPHET and 3R have better performance than Epidemic and Spray and Wait when the memory size is small (<5MB). Since the MIT Reality trace lasts for 1 month and these two routings can make use of the contact history to get more accurate prediction for message forwarding, a smaller memory is relatively enough due to less replicas and message forwarding. As the memory size increases, Epidemic and Spray and Wait achieve higher delivery rate because a large number of flooding replicas can be tolerated in the memory with few of message dumps.

Similarly, the message overhead (MO) in Fig. 4(b) and communication overhead in Fig. 4(c) vary in the same way as the memory size increases. Since there are more nodes in the MIT Reality data trace than in INFOCOM trace, the overheads of Epidemic are significantly increased. That means Epidemic is not suitable for DTNs with a large number of devices. We can see that the overheads of Spray and Wait increase with the increase of maximum allowed replicas (from 5 in INFOCOM trace to 70 in MIT Reality trace). PRoPHET and 3R protocols have lower overhead for both data traces, which is determined by the popularity of nodes and the activeness of each node.


From this set of simulations, it can be seen that flooding-based routing protocols can achieve better delivery rate, but require much larger message memory. Whereas, the prediction-based routing protocols are more memory efficient, and can achieve desirable delivery rate with very low overheads as they have an accurate prediction on the contact patterns.

4.3. Performance Evaluation with Different Packet Lifetimes

Each message may experience an unpredictable delivery latency in DTNs due to the lack of guaranteed continuous end-to-end connectivity. Therefore, in order to achieve more effective communication for DTNs, each message needs a suitable lifetime either by users setting or default configuration of specific devices.In this set of simulations, we evaluate the influence of different packet lifetimes for the routing algorithms using INFOCOM and MIT Reality data traces respectively. We assume each node has an infinite buffer size. (1) INFOCOM Trace:Fig. 5 shows the delivery rate and overheads of four routing protocols with different setting of packet lifetimes using the INFOCOM trace. In Fig. 5(a), it can be seen that for each routing protocol the delivery rate appears with a dramatic increase when the packet lifetime increases from 6 to 30 hours. For example, the delivery rate increases from 0.454 to 0.832 for 3R routing. The reason is that the probability of directly encountering the destination or the message delivered to the destination by multi-hop forwarding grows when the packet lifetime increases. This increasing trend slows down when the packet lifetime exceeds 30 hours. That is because the data trace only lasts 4 days, and the contact pattern is short-time dependent. The delivery rate will not increase even with a large packet lifetime since the destination node may only be active in a short time period. The overheads shown in Fig. 5(b) and (c) also demonstrate that, for Epidemic and PRoPHET, the overheads increase as the packet lifetime increases, since more replicas can be generated and forwarded to the destination. Whereas, the lifetime is enough long for message delivery when it exceeds 30 hours in this trace. The overheads of Spray and Wait and 3R keep stable with different packet lifetime, because the number of replicas is limited in Spray and Wait and 3R is a single-message forwarding based scheme.

Fig.5.Performance evaluation with different packet lifetimes using INFOCOM trace, (a) delivery rate, (b)

message overhead, (c) communication overhead.


Fig.6.Performance evaluation with different packet lifetimes using MIT Reality trace, (a) delivery rate, (b) message overhead, (c) communication overhead.

(2) MIT Reality Trace: The simulation results with MIT Reality trace are similar as shown in Fig. 6. It can be seen in Fig.6(a) that the delivery rate for all the routing protocols keep growing as the packet lifetime increases. It is because MIT Reality trace lasts for 1 month and it is long-time dependent, thus a longer packet lifetime always provides a larger probability of successful delivery. In this simulation, as shown in Fig.6(b) and (c) the overheads of Epidemic increase dramatically both due to longer packet lifetime and larger device population. Whereas, the overheads variations for other routing protocols are similar to the results using INFOCOM data trace. This set of simulations demonstrate that a longer packet lifetime can provide a higher delivery rate for DTNs, since there is more chance to encounter the destination node either by direct contact or multi-hop message forwarding. The overheads of flooding-based routings increase significantly, while the prediction-based routings are more efficient, especially when the DTNs have long-time dependent contact patterns.

5. CONCLUSION

Due to the property of Delay Tolerant Networks for no guaranteed continuous connectivity, choosing a proper routing protocol and finding its optimal configuration are challenging problems. In this paper, we evaluate some existing representative routing protocols for DTNs through extensive trace-based simulations. Flooding-based routing protocols, such as Epidemic and Spray and Wait, can achieve better delivery rate by sacrificing the memory, while prediction-based routings, like PRoPHET and 3R, are more efficient when considering the delivery overheads. This paper provides important guidances for routing protocols design. Our future work is to design an efficient routing protocol for DTNs with high delivery rate and low delivery overheads.

REFERENCES

[1] T. Cao, X. Zhang, L. Kong, X. Liu, W. Shu, and M. Wu, “Traffic awarerouting in urban vehicular

networks,” in WCNC, pp. 2004–2009, 2013. [2] P. Juang, H. Oki, Y. Wang, M. Martonosi, L. S. Peh, and D. Rubenstein,“Energy-efficient computing

for wildlife tracking: Design tradeoffs andearly experiences with zebranet,” SIGARCH Comput. Archit. News,vol. 30, no. 5, pp. 96–107, 2002.


[3] Y. Zhang and J. Zhao, “Social network analysis on data diffusion indelay tolerant networks,” in MobiHoc, pp. 345–346, 2009.

[4] A. Vahdat and D. Becker, “Epidemic routing for partially-connected adhoc networks,” tech.rep., 2000.

[5] T. Spyropoulos, K. Psounis, and C. S. Raghavendra, “Spray andwait: An efficient routing scheme for intermittently connected mobilenetworks,” in WDTN, pp. 252–259, 2005.

[6] A. Lindgren, A. Doria, and O. Schel´en, “Probabilistic routing in intermittentlyconnected networks,” SIGMOBILE Mob.Comput.Commun.Rev., vol. 7, pp. 19–20, July 2003.

[7] J. Burgess, B. Gallagher, D. Jensen, and B. Levine, “Maxprop: Routingfor vehicle-based disruption-tolerant networks,” in INFOCOM, pp. 1–11, April 2006.

[8] Q. Yuan, I. Cardei, and J. Wu, “An efficient prediction-based routingin disruption-tolerant networks,” IEEE Trans. Parallel Distrib.Syst,vol. 23, no. 1, pp. 19–31, 2012.

[9] L. Vu, Q. Do, and K. Nahrstedt, “3r: Fine-grained encounter-basedrouting in delay tolerant networks,” in WoWMoM, pp. 1–6, 2011.

[10] A. Lindgren, A. Doria, E. Davies, and S. Grasic, “Probabilistic routingprotocol for intermittently connected networks,” Internet-Draft, 2012.

[11] J. Scott, R. Gass, J. Crowcroft, P. Hui, C. Diot, and A. Chaintreau,“CRAWDAD data set cambridge/haggle (v. 2006-01-31).”http://crawdad.org/cambridge/haggle/, 2006.

[12] N. Eagle and A. (Sandy) Pentland, “Reality mining: sensing complexsocial systems,” Pers. Ubiq.Comput., vol. 10, no. 4, pp. 255–268, 2006.

[13] D. Kotz, T. Henderson, and I. Abyzov, “CRAWDAD data set dartmouth/campus (v. 2004-12-18).” http://crawdad.org/dartmouth/campus,2004.

AUTHORS

Luming Wan, is a Master candidate from Department of Computer Science, University of Otago, New Zealand. He obtained B.S. degree from University of Otago, New Zealand in 2012. His research interests include Delay Tolerant Network (DTN), Social Network, etc. Feiyang Liu,is a PhD candidate from Department of Computer Science, University of Otago, New Zealand. He obtained B.S. and M.S. degrees of telecommunication engineering from Xidian University, China in 2009 and 2012, respectively. His research interests include Network on Chip (NoC), Optical Network on Chip (ONoC), Delay Tolerant Networks (DTN), etc. Juan Zhang, received B.S degree in 2011from Jiangsu University,China, and M.S. degree of control science and engineering in 2014 from Beihang University, China. Her research interests include distributed algorithms and protocols, wireless sensor networks, cyber-physical systems, and data communication, etc. Haibo Zhang, received the MSc degree in Computer Science from Shandong Normal University, China in 2005, and the PhD degree in Computer Science from the University of Adelaide, Australia in 2009. From 2009 to 2010, he was a postdoctoral research associate at Automatic Control Laboratory, KTH, Sweden. Currently he is a lecturer at Computer Science department of University of Otago, New Zealand. His research interests include real-time industrial wireless communications, wireless sensor/ad hoc networks, delay-tolerant networks, green computing, distributed algorithms and protocol design.

David C. Wyld et al. (Eds) : WiMo, ITCSE, ICAIT, NC - 2015

pp. 13–25, 2015. © CS & IT-CSCP 2015 DOI : 10.5121/csit.2015.51002

MODELING SOCIAL GAUSS-MARKOV

MOBILITY FOR OPPORTUNISTIC

NETWORK

GuoDong Kang1 Jinzhi Ning

1 and GuoLiang Kang

2

1DFH Satellite Co., Ltd., 100094, Beijing, China

[email protected] 2University of Technology, Sydney 15 Broadway, Ultimo NSW 2007

[email protected]

ABSTRACT

Mobility is attracting more and more interests due to its importance for data forwarding

mechanisms in many networks such as mobile opportunistic network. In everyday life mobile

nodes are often carried by human. Thus, mobile nodes’ mobility pattern is inevitable affected by

human social character. This paper presents a novel mobility model (HNGM) which combines

social character and Gauss-Markov process together. The performance analysis on this

mobility model is given and one famous and widely used mobility model (RWP) is chosen to

make comparison..

KEYWORDS

Mobility, Social, Gauss-markov, Opportunistic

1. INTRODUCTION

With rapid growing popularity of small portable devices (e.g., smart phones, PDAs, laptops) in

personal communication domain, a prospective networking scenario occurs where a large number

of mobile nodes share or forward information by means of opportunistic encounter[16]. Since

these mobile nodes are carried by human, modeling mobility based on realistic human mobile

pattern is becoming an essential and important thing. Normally people live in various social

circles such as their family circle, their friends circle and so on. This kind of social property

influences human movement pattern and further affects nodes’ pattern as well. Sometimes in a

social circle some people are more popular than others and have more chance to encounter others.

The nodes carried by this kind of person can be regarded as one suitable choice for relaying data

with other social circles. In other words, these persons could be selected as heads of the

corresponding circles and the nodes carried by them could be regarded as one head node. To

construct this kind of social opportunistic mobility, this paper adopts random graph for setting up

social relations and Gauss-Markov theory for solving individual mobile pattern.


The rest of this paper is organized as follows: Section II describes the allocation method of

mobile nodes. Section III proposes the mathematical model of the mobility. Section IV gives

analysis of the proposed mobility model. Conclusion is given in Section V.

2. GROUP ALLOCATION METHOD

People’s social circle character has a significant impact on nodes mobility pattern. Nodes in

different social circles can be allocated into different groups. Node often has more relations with

other nodes inside the same group or circle than with those nodes belonging to different groups or

circles.

2.1. Node Relation Setup

Normally, there is some kind of social or biological relations among those nodes since they are

carried by people. To set up this kind of mutual relation, we adopt the classical method of

representing social or biological network, weighted graphs. The strength of mutual relation

between any node pair is represented using a value in the range [0, 1]. As a consequence, the

network internal relation can be described as a relation matrix with a dimension of N*N where N

is the total number of nodes in the network. At present, there are several models that describe the

key properties of real-life network, such as random graph, small world, etc. Some research work

show that the properties of these random graphs, such as path lengths, clustering coefficients,

can’t be regarded as accurate models of realistic networks [5, 6, 7]. Here, we choose the

geometric random graph to be the network model. In this kind of model, the geometry relations

of nodes have strong association with the social relation of nodes. That means when any two

nodes are in the radio range of each other, the social relation exists. On the contrary, since they

even can’t communicate with each other, we think that there is no any social interaction between

them. So when the Euclidean distance between any two nodes is smaller than radio range R, the

corresponding element of the social relation matrix M is set to be 1 or else set to be 0 as follows.

,

1 ,

0

i j

i j

if i j and P P Rm

otherwise

≠ − ≤=

It shall be emphasized that the relations value of one node and itself is regarded to be zero in the

matrix. In [6] it is shown that in two or three dimensional area using Euclidean norm can supply

surprisingly accurate reproductions of many features of real biological networks. Fig.1 gives one

example of 100 nodes.

2.2. Nodes Group Allocation

Once the relation matrix M is obtained, groups can be detected. Group structure is one of the

common characters in many real networks. However, finding group structures within an arbitrary

network is an acknowledged difficult task. A lot of work has been done on that. Currently, there

are several methods that can achieve that goal, such as Minimum-cut method, Hierarchical

clustering, Girvan-Newman algorithm, etc.


Fig.1. An example of 100 Nodes Relations. (Small circles represent nodes carried by people which are

uniformly distributed in one given 100m*100m square; the short straight lines represent the existing

relation between node pair.)

Minimum-cut method is one of the oldest algorithms for dividing networks into parts. This

method uses in load balancing for parallel computing in order to minimize communication

between processor nodes. However, this method always finds communities regardless of whether

they are implicit in the structure, and it can only find a fixed number of them. So it is less than

ideal for finding community structure in general networks [4].

Hierarchical clustering is another method for finding community structures in network. This

method detects the community by defining a similarity measure quantifying some (usually

topological) type of similarity between node pairs.

The Girvan–Newman algorithm is one commonly used algorithm for finding communities [12]. It

identifies edges in one network that lie between communities and then removes them, leaving

behind just the communities themselves. But this method runs slowly which makes it impractical

for networks of more than a few thousand nodes [11].

Modularity maximization is one of the most widely used methods for community detection [11].

Modularity is a benefit function that measures the quality of a particular division of a network

into communities. This method detects the community structure of high modularity value by

exhaustively searching over all possible divisions [8].

In this paper we adopt modularity maximization as the social group detection method. Modularity

maximization is one of the most widely used methods for group detection [11].


into groups. This method detects the group structure of high modularity value by exhaustively

searching over all possible divisions [8].


In real networks, the modularity value is usually in the range [0.3, 0.7]; 1 means a very strong

group structure and 0 means no better than random.

3. MOBILITY MODEL SETUP

3.1. Mobility Scenario Description

One kind of usual mobility scenario appearing in daily life is one people in a group leading the

mobility. This person is selected because of his popularity in the group or his familiarity with the

environment. He is called as head of the group in this paper. The head moves freely and leads

other people to move together. Other people follow head people’s direction and adjust their

speeds to keep up with him. The head’s direction and speed determine the general direction and

speed of whole group. Fig.2 gives an illustration of this mobility scenario.

Head Node

Head Node

Follow

Node 2

Follow

Node 2

Follow

Node 1

Follow

Node 1

Group Direction

Fig.2. Illustration of mobility scenario. The red solid circle represents head node or head people. The other

two blue solid circles represent follow nodes or follow people. The dash circles represent social circles or

groups which head node and follow nodes belong to.

3.2. Mobility Mathematical Model

Assume there are total N nodes in one group, the th

n node is denoted as node ( )1n n N≤ ≤

3.2.1 Head Node Determination Scheme

Assume at instant t there are number M nodes which react with node i. The following expression

describes the average relation value of node i with other nodes :

,

1

1( ) ( ) ( )

N

i i j

j

m t m t i jM =

= ≠∑


where

N is the total number of the network;

M is the number of nodes in the network characterized by ,

0i j

m >

( )i

m t is the average relation value of node i at instant t;

, ( )i jm t is the element of the relation matrix at instant t;

In this paper, the node which has the greatest average relation values will be assigned to be the

Head node of the group and the remaining nodes in the group will be regarded as Follow nodes

which will follow the head node. At the beginning, the head node of the group is denoted as node

h, it has

| |( 0) ( 0) (1 , )h i

m t t m t t i N i h= ≥ = ≤ ≤ ≠

where ( )h

m t is the average relation value of node h at time t;

Initially, the head node is chosen as the node which is most popular in the group, namely has the

strongest relations with other member nodes of the group. However, as the time past, the mutual

relation between the initial assigned head node and other nodes in the group may vary. Maybe at

time e

t the initial head node don’t have the strongest relation with others any more. Consequently,

one election shall take place to determine the next new head node who has the strongest relation

with other nodes at time e

t . The time e

t is called head node determination period in this paper.

The average relation value of the new elected head node, E

h , has

( ) ( ) (1 , )Eh e i eE E Nm t t t m t t t i i h• •= ≥ = ≤ ≤ ≠

where

0E ≥ is the head node election number in the group;

et is the head node election period.

3.2.2. Gauss-Markov process setup

A Gauss-Markov process is a simple stochastic process with many social applications. In terms of

continuous time, a stationary Gauss-Markov process can be described by the following

autocorrelation function [16,18]:

[ ] 2( ) ( ) ( )

fR E f t f t e

β ττ τ σ −= + = (1)

where

2

σ is the variance;

0β ≥ determine the degree of memory in the process.

Assume to discretize time into serial time slots t∆ (normalized to 1 throughout this paper),

( ) ( )f t f s t= ⋅ ∆


Define t

eβ

α− ∆

= , then the discrete representation of (1) is

[ ] 2

( ) ( 1) (1 ) 1Gauss

s t s t df fα α µ α⋅ ∆ = − ⋅ ∆ + − + −

where

0 1α≤ ≤ ;

µ is the asymptotic mean of f

when s approach infinity.

Gaussd is an independent, uncorrelated, and stationary Gaussian process, with mean

Gaussµ equals

to zero and standard deviation Gauss

σ equals to σ when s approaches infinity.

3.2.3. Gauss-Markov mobility discussion

In this paper, nodes’ speed and direction are both regarded as Gauss-Markov processes which are

correlated with time. Nodes’ speed and direction discrete representations of (1) can be described

as follows:

[ ] 2( ) ( 1) (1 ) 1

v

v GaussV s t V s t dα α µ α⋅ ∆ = − ⋅ ∆ + − + −

[ ] 2( ) ( 1) (1 ) 1

Gausss t s t d

φ

φα α µ αΦ ⋅ ∆ = Φ − ⋅ ∆ + − + −

Gauss-Markov allows previous speed or direction to influence current speed or direction.

Moreover, Gauss-Markov is a memory-tunable process accounting for its suitableness for the

description of relations of follow nodes and head node about speed and direction choices. The

parameter, α , is tunable which determines the amount of memory and variability in movement.

The following discusses three cases of parameter α .

Case 1: 0α =

When 0α = , Gauss-Markov becomes memoryless; The speed and direction of nodes completely

depend on their average v

µ ,φ

µ and the Gauss variables.

( )v

v GaussV s t dµ⋅ ∆ = +

( )Gauss

s t dφ

φµΦ ⋅ ∆ = +

This paper thinks this case is quite suitable for setting up the relations of speed and direction

between follow nodes and head node by setting fnv

µ ,fnφ

µ of follow nodes to identify hv

µ ,hφ

µ of

head node since follow nodes have one common character that is they always try to keep up with

head node movement.

Case 2: 1α =

When 1α = , Gauss-Markov loses all the variables becoming predictable pattern. The current

speed and direction is totally identical to the previous values. The movement becomes a straight

line.


[ ]( ) ( 1)V s t V s t⋅ ∆ = − ⋅ ∆

[ ]( ) ( 1)s t s tΦ ⋅ ∆ = Φ − ⋅ ∆

This paper thinks this case is suitable for head node mobility since his destination is explicit.

Case 3: 0 1α< <

This is intermediate randomness case between the two above cases. This case won’t be taken into

account too much here.

3.2.4 Head node mobility setup

Normally, the head node in a group shall know its destination explicitly. Once he determines the

destination, he will move straightly toward the destination maintaining his speed and direction

until he reaches the destination. During head node mobility, he may have several destinations

thus has several corresponding mobility sections. Each mobility section is one 1α = case Gauss-

Markov process. The speed and direction of different mobility sections are different. Head node

mobility is set up as follows.

Assume at instant i

t one destination of head node h is chosen and at instance jt the head node

reaches this destination. The destination of head node can be denoted as

[ ]( ) ( ) ( )dh i j dh i dh i

P t t t t X t Y t≤ ≤ =

Normalize t∆ to be 1, at instant t ( ,i j

t t t s tt≤ ⋅ ∆< = ), the position of node h is a function of

coordinate ( )h

X t and ( )h

Y t which can be expressed as follows:

( ) cos ( )

( ) sin ( )

( ) 1

( ) 1

( )( )

( )

h h

h h

h h

h h

h

X XP

Y

V s t s t

V s t s t

s t s t

s t s ts t

Y

⋅ ∆ ⋅ Φ ⋅ ∆

⋅ ∆ ⋅ Φ ⋅ ∆

⋅ ∆ − ⋅ ∆

⋅ ∆ − ⋅ ∆

+⋅ ∆ = =

+

where

s is the number of t∆ at current instant;

1s − is the number of t∆ at previous instant;

h

Φ is the direction function of node h which can be defined as

( ) ( )( )

( ) ( )

dh i h i

h

dh i h i

Y t Y tt arctg

X t X t

−Φ =

−

hV is the speed function obeying normal distribution which is determined with destination

synchronously.

The research in [13] has shown that the walking speed of a human obeys normal distribution. The

measurement in [14] shows that the mean speed of a walking human is a range which is from

1.16 m/s to 1.58 m/s representing walking normally or walking fast. In [15], the Manual of

Uniform Traffic Control Devices (MUTCD) shows that human walk with a normal speed of 1.2

m/s (4 ft/sec). [10] indicates a statistics that walking speed for younger human (ages 14 to 64)


was 1.25 m/sec (4.09 ft/sec); for older human (ages 65 and over) it was 0.97 m/sec (3.19 ft/sec).

For designing purposes values of 1.22 m/sec (4 ft/sec) for younger human and 0.91 m/sec (3

ft/sec) for older human are appropriate [10]. In this paper, the mean value of walking speed of

the head node, hv

µ , equals to 1.22 m/s [10] and its standard deviation hv

σ is 0.26 m/s[13]. So the

head node’s speed at time t is

( ) ( ) ~ ( , )h h i hv hv

V t V t N µ σ=

The probability density function of ( )h

V t can be expressed as:

( )2

2exp

2( )

2

hv

hv

hv

v

f v

µ

σ

σ π

−− =

Once head node reaches the destination, he will go towards his next destination which is

uniformly chosen in given movement area.

3.2.5 Follow nodes mobility setup

Follow nodes’ speed and direction discrete representation of (1) can be described as follows

according to case 1:

( )v

fn fnv GaussV s t dµ⋅ ∆ = +

( )fn fn Gauss

s t dφ

φµΦ ⋅ ∆ = +

where

fnvµ equals to

hvµ ;

v

Gaussd is an independent, uncorrelated, and stationary Gaussian process, with mean

v

Gaussµ equals

to 0 and standard deviation v

Gaussσ equals to

hvσ .

Similarly, this paper chooses fn hφ

µ = Φ and /12.Gauss

φσ π=

Consequently, the position of follow node fn is a function of coordinate ( )fn

X s t⋅ ∆ and

( )fn

Y s t⋅ ∆ which can be expressed as follows:

( ) cos ( )

( ) sin ( )

( ) 1

( ) 1

( )( )

( )

fn fn

fn fn

fn fn

fn fn

fn

X XP

Y

V s t s t

V s t s t

s t s t

s t s ts t

Y

⋅ ∆ ⋅ Φ ⋅ ∆

⋅ ∆ ⋅ Φ ⋅ ∆

⋅ ∆ − ⋅ ∆ ⋅ ∆ − ⋅ ∆

+⋅ ∆ = =

+

The mean of speed and direction of follow nodes equaling to the values of head node interprets

that their mobility shall follow group’s movement. The difference of standard deviation of

normal distribution between follow-node and head-node reflects these follow-nodes have their

freedom degree in mobility to some extent.


4. MODEL CHARACTER ANALYSIS

4.1 Classical Individual Mobility Model-RWP

0 0( , )x y

1 1( , )x y

2 2( , )x y

3 3( , )x y

4 4( , )x y

5 5( , )x y

6 6( , )x y

7 7( , )x y

8 8( , )x y

Fig.3. An illustration of RWP

Josh et al. [8] present a Random Waypoint model. In this model a mobile node is initially placed

at random location in the simulation area with a destination and speed according to the random

distribution. Then the mobile node moves from its current location to its destinations via a

straight line. After arrival at the destination, the mobile node determines a pause time according

to a random distribution, and after that chooses a new destination.

4.2 Analysis and Comparision

To analyze the character of these two mobility models, two important parameters inter-contact

time and contact time are adopted. These two parameters describe the characteristics of

connection opportunities of the network, i.e. how many and when they occur, how often and how

long. Contact time is defined as the time interval during which the two nodes can keep contact.

The time interval from this contact to next one is defined as inter-contact time during which

nodes can’t communicate [1]. Fig.4. gives an illustration. These two parameters are very

important for opportunistic network. Contact time can help us determining the capacity of

opportunistic networks. Inter-contact time is a parameter which strongly affects the feasibility of

the opportunistic networking.

The new mobility model in this paper is named as Head Node Group Mobility Model (HNGM)

here. We let each mobility model run 1000 seconds for one experiment and made 50 times

similar experiments for each mobility model. Head node election period is set to be 1000 seconds

here. Figure 5 and Figure 6 give the contact time distribution and inter-contact time distribution

in different coordinate systems.


Fig.4. Contact and Inter-contact times for a pair of nodes[1].

In Fig. 5 (a), we can see that the new mobility model’ inter-contact time distribution behave an

exponential distribution plot using log-log coordinate system. In Fig. 5 (b), the contact time

distributions show more evident difference. The RWP mobility models’ curves are still

exponential-like curves. In contrast, the shape of HNGM model’ curves are changed. The HNGM

model’s curve is one kind of transition from exponential-like curve to power- law-like curve.

This shows HNGM model behaves stronger group character than RWP model. In Fig. 6 we can

see the difference between the curves more clearly using semi-log coordinate system. The

exponential nature of the inter-contact time shown in Fig. 5 (a) becomes straight line form in Fig.

6 (a).

Fig5 (a)

100

101

102

103

10-3

10-2

10-1

100

time (second)

P[X

>x]

Inter-contact time cdf

RWP

HNGM


Fig5 (b)

Fig. 5. Inter-contact time cdf and contact time cdf in log-log coordinates

Fig6 (a)

100

101

102

103

10-3

10-2

10-1

100

time (second)

P[X

>x]

Contact time cdf

RWP

HNGM

0 200 400 600 800 100010

-3

10-2

10-1

100

time (second)

P[X

>x]


RWP

HNGM


Fig 6 (b)

Fig. 6. Inter-contact time cdf and contact time cdf in semi-log coordinates.

5. CONCLUSION AND FUTURE WORK

This paper proposed one new group mobility model using social Gauss-Markov scheme which is

based on one social realistic mobility scenario. This paper also analyzes its characters about

contact time, inter-contact time. The curves of contact time and inter-contact time show different

shapes with the classical individual model RWP. By comparison, HNGM model shows obvious

group character than RWP model. In the future work, the author will undertake further research

on mobility about 0 1α< < case of Gauss-Markov process.

ACKNOWLEDGEMENT

The author would like to thanks DFH Corporation support.

REFERENCES

[1] A. Chaintreau, P. Hui, J. Crowcroft, C. Diot, R. Gass, and J. Scott, “Pocket Switched Networks:

Real-world mobility and its consequences for opportunistic forwarding,” Technical Report UCAM-

CL-TR-617, University of Cambridge, Computer Laboratory, February 2005.

[2] J Broch, DA Maltz, DB Johnson, YC Hu, J Jetcheva,” Multi-hop wireless ad hoc network routing

protocols”, in Proceedings of the ACM/IEEE International Conference on Mobile Computing and

Networking (Mobicom 1998), pages 85-97, 1998.

[3] Y. Ko and N.H. Vaidya, “Location-aided routing (LAR) in mobile ad hoc networks,” in Proceedings

of the ACM/IEEE International Conference on Mobile Computing and Networking (Mobicom),

pages 66-75, 1998.

[4] M. E. J. Newman, Detecting community structure in networks. Eur. Phys. J. B, 38, 321–330 (2004)..

0 200 400 600 800 100010

-3

10-2

10-1

100

time (second)

P[X

>x]

Contact time cdf

RWP

HNGM


[5] E. de Silva and M. Stumpf, “Complex networks and simple models in Biology”, J. R. Soc. Interface,

2 (2005), pp. 419-430..

[6] N. Przulj, D. G. Corneil, and I. Jurisica, “Modeling interactome:Scale- free or geometric?”,

Bioinformatics, 20 (2004), pp. 3508-3515..

[7] D. J. Watts and S. H. Strogatz, “Collective dynamics of 'small-world' networks”, Nature, 393 (1998),

pp. 440-442.

[8] M.E.J. Newman and M. Girvan. “Finding and evaluaing community structure in networks”. Physical

Review E,68,2003.

[9] Mirco Musolesi, Cecilia Mascolo, “A community based mobility model for ad hoc network research”,

in Proceedings of the 2nd international Workshop on Multi-Hop Ad Hoc Networks: From theory To

Reality (REALMAN '06), Florence, Italy, May 2006, pp 31-38.

[10] R.L. Knoblauch, M.T. Pietrucha, M. Nitzburg. Field studies of pedestrian walking speed and start-up

time. Transportation Research Board Records No. 1538, 1996.

[11] Newman, M. E. J. Fast algorithm for detecting community structure in networks. Phys. Rev. E

69,066133 (2004)

[12] Girvan, M. and Newman, M. E. J., Community structure in social and biological networks, Proc.

Natl. Acad.Sci. USA 99, 8271–8276 (2002).

[13] L. Henderson. The statistics of crowd fluids. Nature, Vol. no. 229, 381-383, 1971.

[14] Finnis, K.K. and Walton, D. Field observations of factors influencing walking speeds. Ergonomics,

2006.

[15] LaPlante, John and Kaeser, Thomas P. “A History of Pedestrian Signal Walking Speed

Assumptions”, in Proceedings of the 3rd Urban Street Symposium: Uptown, Downtown, or Small

Town: Designing Urban Streets That Work, 2007.

[16] L. Pelusi, A. Passarella, and M. Conti, “Opportunistic networking: data forwarding in disconnected

mobile ad hoc networks,” Communications Magazine, IEEE, vol. 44, no. 11, pp. 134 –141, november

2006.

[17] B. Liang and Z. Haas. Predictive distance-based mobility management for PCS networks. In

Proceedings of the Joint Conference of the IEEE Computer and Communications Societies

(INFOCOM), March 1999

[18] E. Wong and B. Hajek, Stochastic Processes in Engineering Systems,Springer-Verlag, 1985

AUTHORS

GuoDong KANG is an telecommunication engineer of DFH Satellite Co., Ltd., 100094,

Beijing, China.


INTENTIONAL BLANK



A TRIANGLE-TRIANGLE INTERSECTION

ALGORITHM

Chaman L. Sabharwal and Jennifer L. Leopold

Missouri University of Science and Technology

Rolla, Missouri, USA – 65409 {Chaman,leopoldj}@mst.edu

ABSTRACT

The intersection between 3D objects plays a prominent role in spatial reasoning, geometric

modeling and computer vision. Detection of possible intersection between objects can be based

on the objects’ triangulated boundaries, leading to computing triangle-triangle intersection.

Traditionally there are separate algorithms for cross intersection and coplanar intersection.

There is no single algorithm that can intersect both types of triangles without resorting to

special cases. Herein we present a complete design and implementation of a single algorithm

independent of the type of intersection. Additionally, this algorithm first detects, then intersects

and classifies the intersections using barycentric coordinates. This work is directly applicable to

(1) Mobile Network Computing and Spatial Reasoning, and (2) CAD/CAM geometric modeling

where curves of intersection between a pair of surfaces is required for numerical control (NC)

machines. Three experiments of the algorithm implementation are presented as a proof this

feasibility.

KEYWORDS Intersection Detection, Geometric Modeling, Classification Predicates, Spatial Reasoning,

Triangle-Triangle Intersection.

1. INTRODUCTION

The intersection between 3D objects plays a prominent role in several areas including spatial

reasoning[1], real-time rendering [2], collision-detection, geology [3], geometric modeling

[4],and computer vision. Traditionally there are specialized algorithms for cross intersection and

coplanar intersection. The intersection detection is a byproduct of actual intersection

computations. Most of the time intersection detection should be done before the actual

intersection. For example, in qualitative spatial reasoning, intersection detection is sufficient,

actual intersection is not even needed, whereas in geometric applications like surface-surface

intersection, it is desirable to have actual intersections. For early detection of intersection, we

present a complete uniform algorithm independent of whether the triangles have cross

intersection or coplanar intersection. Herein we illustrate this with an algorithm, its Python

implementation is supported with output displayed in tables and figures. Typically, the boundary

of a3D object is represented as a triangulated surface and a triangle-triangle intersection is the

computational basis for determining intersection between objects. Since an object boundary may


contain thousands of triangles, algorithms to speed up the intersection detection process are still

being explored for various applications, sometimes with a focus on innovations in processor

architecture[5].

In qualitative spatial reasoning, spatial relations between regions are defined axiomatically using

first order logic [6] or the 9-Intersection model [1]. It has been shown in [7] that it is sufficient to

define the spatial relations by computing 4-Intersection predicates, (namely, Interior–Interior

(IntInt), Boundary–Boundary (BndBnd), Interior–Boundary (IntBnd), and Boundary–Interior

(BndInt)) instead of 9-Intersections. In order to implement these algorithms, we must first

implement the triangle-triangle intersection determination.

In geometric modeling, the surface-surface intersection problem occurs frequently. Since a free-

form surface is represented as triangulated mesh, intersection between 3D objects amounts to

detecting and computing intersection between 3D triangles. The algorithm presented here is

extremely useful for geometric modeling where intersection between objects occurs thousands of

times. For geometric applications cross intersection is most often used to the determine curves of

intersection between surfaces [4]. Most of the approaches to these problems are based on code

optimization of same or similar cross intersection algorithms. No attempt is made to design a

single algorithm that handles both intersections simultaneously. Herein we present a new

algorithm that fills the gap.

This paper is organized as follows: Section 2 discusses the background of possible cross and

coplanar intersection between a pair of triangles. It describes the cross intersection, coplanar area

intersection algorithm, and composite algorithms for general triangles. Section 3 develops the

new generic single algorithm for triangle-triangle intersection, and classifies the intersections.

Section 4 describes implementation experiments and the timing results. Section 5describes the

two of the applications where this approach is directly applicable. Section 6concludes and

references are given in Section 7.

2. BACKGROUND: TRIANGLE-TRIANGLE INTERSECTION

There is an abundance of papers devoted to intersection between a pair of triangles

[8,9,10,11,12]. Interestingly, most of the papers concentrate on cross intersection, they simply

reinvent the algorithm and optimize the code to implement it slightly differently and more

efficiently, with no innovation. The cross intersection may result in a no intersection, a single

point intersection or a line intersection. For coplanar triangles, it can result in area intersection as

well. For coplanar intersection, such cross intersection algorithms do not work, there are separate

algorithms for area intersection [10,11,12]. In these approaches, they first try to compute the

intersection. Whether intersection is found or not, then it leads to conclude on the existence of

intersection. The paper [12]surveyed various approaches for determining the cross intersection

detection, and developed a fast vector version of the cross intersection detection, as well as

classification of the type of intersection. The papers [8] and [11] also compare hardware

implementation of their algorithm based on the number of arithmetic operations: +, -, *, /.

Another paper [10] also compares the optimized intersection times. The papers[10,11] considered

an approach for determining the intersection detection covering coplanar intersection, using edge

to edge intersections. These approaches [10,11,12] led to two separate algorithms one for cross

and one for coplanar intersection. There is no single algorithm that can handle both cross and

coplanar intersection. We present a new algorithm that is analytically different, exhaustive and


more rigorous than the previous algorithms [8,9,10,11,12]. It computes intersection using

barycentric coordinates and vector equations very judiciously. Logical rather than computational

tests are used to detect whether the intersection is a single point, or a line or an area. This new

algorithm encompasses all cases implicitly and is different from previous ones in that we use only

one equation rather than different algorithm for each special case. We show that it is possible to

detect the existence of intersection in a uniform way before the precise intersection computations

are computed.

2.1 Specialized Intersection Methods and Algorithms

Here we describe the conventional approach to triangle-triangle intersection in terms of two

algorithms, one for cross intersection and one for coplanar intersection. Then we derive the

composite algorithm for triangle-triangle intersection. Here we give the line intersection for cross

intersection and area intersection for coplanar triangles separately and combine the two into one

algorithm as is conventionally done. Then in Section 3 we present our new algorithm, which is a

single algorithm that employs a more singular, seamless approach.

2.1.1 The Triangle-Triangle Line Intersection Algorithm

The cross intersection algorithm encompasses the single point and edge intersection cases. Here

we give a solution different from previous approaches.

Algorithm for cross intersection line Algorithm

Input: Two triangles ABC and PQR

Output: Determine if they cross intersect. Return true if they intersect, otherwise return false.

Boolean crossInt (tr1 = ABC, tr2 = PQR)

The vector equations for two triangles ABC and PQR are

R1(u, v) = A + u U + v V, 0 ≤ u, v, u + v ≤ 1

R2(s, t) = P + s S + t T, 0 ≤ s, t, s + t ≤ 1

where U = B-A, V = C - A, and S = Q-P, T = R – P are direction vectors along the sides of

triangles.

Let N1 =UxV, N2 =SxT be normals to the planes supporting the triangles. The triangles intersect

if there exist some barycentric coordinates (u, v) and (s, t) satisfying the equation

A + u U + v V = P + s S + t T

Since N1xN2 ≠ 0 for cross intersecting triangles, and S and T are orthogonal to N2, the dot product

of this equation with N2 eliminates S and T from the above equation to yield

u U•N2 + v V•N2 = AP•N2

This is the familiar equation of a line in the uv - plane for real variables u, v. The vector equation

using real parameter λ becomes

If there is a λ satisfying three equations such that 0 ≤ u, v, u + v ≤ 1, then the triangles are

ensured to intersect. The solution in λ is the range of values λm ≤ λ ≤λM for some λm and λM.

(u,v) = AP • N2

(U • N2,V • N2 )

U • N2

2 +V • N2

2+ λ(V • N2,−U • N2 )


This gives the line of intersection of uv-parameter triangle with the st-parameter plane. Similarly

the line of intersection of st-triangle with the uv-plane is computed. Then the common segment if

any is the line intersection between the two triangles, for details see [9,13].This algorithm works

only if the triangles cross intersect.

2.1.2 The Triangle-Triangle Area Intersection Algorithm

The area intersection is possible for coplanar triangles only. If the triangles ABC and PQR are

coplanar and a vertex of PQR is in the interior of ABC (or the converse is true), then an area

intersection occurs. If there is no edge-edge intersection and no vertex of one triangle is inside the

other triangle (or the converse is true), then they do not intersect, hence they are disjoint. The

input, output, method prototype and pseudocode are given below.

Algorithm for area intersection

Input: Two triangles ABC and PQR coplanar

Output: If they do not intersect, then return false otherwise return true and the intersection area.

Boolean coplanarInt (tr1 = ABC, tr2 = PQR)

The derivation of the algorithm is based on extensive use of the intersections of edges of ABC

triangle with the edges of the PQR triangle and collecting the relevant interactions to form the

area if any, for details see [11]. This algorithm can solve for intersection of coplanar triangles

only.

The classical approach is to use crossInt (tr1, tr2), algorithm from section 2.1.1, when the triangle

cross, and to use the second algorithm coplanarInt (tr1, tr2), algorithm from section 2.1.2, for

coplanar triangles.

Thus separate algorithms are used to determine intersections on case-by-case bases. There is no

single algorithm that detects intersection. We present a new single algorithm that does not depend

on case-by-case intersections.

3. ALGORITHM FOR TRIANGLE-TRIANGLE INTERSECTION

Here we present a single algorithm that is analytically different, more comprehensive, robust and

rigorous; it is implicitly capable of handling any specific type of intersection, which may be no

intersection, a single point, a segment or an area. The Triangles may be coplanar or crossing. This

single algorithm is not a modification of any previous algorithm, it is a new approach different

from the other strategies.

The Main Algorithm Input: Two triangles ABC and PQR in 3D, triangles may be coplanar or crossing

Output: If the triangles do not intersect, return false otherwise return true and the intersection,

which may be single point, a line segment or an area.

Boolean triTriIntersection (tr1 = ABC, tr2 = PQR)

The triangles ABC and PQR are

X = A + u U + v V with U = B - A, V = C - A, 0 ≤ u, v, u + v ≤ 1

X = P + s S + t T with S = Q - P, T = R - P, 0 ≤ s, t, s + t ≤ 1


To detect the intersection between the triangles ABC and PQR, we must attempt to solve

A + u U + v V = P + s S + t Tfor some values 0 ≤ u, v, u + v, s, t, s + t ≤ 1,

Rearranging this equation, we have

uU + v V = AP + s S + t T (1)

whereAP = P - A is a vector

To solve the equation (1) for u,v, we dot equation (1) with (UxV)xV and (UxV)xU, we quickly

get u and v as

u = - ( γ•V + s α•V + t β•V)

v = γ•U + s α•U + t β•U

u + v = γ• (U - V) + s α• (U - V) + t β•(U - V)

where α, β, γ, δ are:

δ = (UxV)•(UxV),

In order to satisfy the constraints 0 ≤ u, v, u + v ≤ 1, we get three inequalities in parameters s

and t

(a) - γ•U ≤ α•U s + β•U t ≤ 1 - γ•U

(b) - 1 - γ•V ≤ α•V s + β•V t ≤ - γ•V

(c) - γ• (U - V)≤α•(U - V) s + β•(U - V) t ≤ 1- γ•(U - V)

inequalities we first eliminate t to get the possible range for s values, then solve for the range

These inequalities (a) - (c) are linear in s and t and are of the general form

m ≤ a s + b t ≤ n

To solve these for possible t values as t(s). If there is no solution, then the algorithm returns

false. If there is a solution, then it return s, t values as

sm ≤ s ≤ sM , tm(s) ≤ t ≤ tM(s).

This discussion is summarized and the intersection points are classified as follows:

if the algorithm returns false,

No Intersection

elseif (sm = sM and (tm(s) = tM(s) for sm ≤ s ≤ sM)

Single Point Intersection

elseif (sm=sMor (tm(s) = tM(s) for sm ≤ s ≤ sM)

Line segment intersection common to two triangles

else

Area Intersection common to two triangles

This completes the discussion of our algorithm for intersection between triangles.

α =S × (U ×V )

δ, β =

T × (U ×V )

δ,γ =

AP × (U ×V )

δ


4. EXPERIMENTAL RESULTS

The algorithm is implemented in

time it utility. Time tests were performed

Core i7). The average time for no intersection,

(1D) and area intersection (2D)

each of 22 sample triangle pairs.

nor hardware embedded, they include classification of intersections also.

proof of concept that this single integrated

data domain consists of synthetic

example, the times for no intersection are averaged over 100

for single point intersection six

used for line segment and averaged over 100 runs, and finally nine sample triangle pairs are used

and averaged. The composite average intersection time is computed. The

displayed in Table 1. We also give

intersection, a line intersection and

figures, see Fig. 1-3. The user interface a

triangle pair for intersections and it displays the corresponding

the sake of space consideration, one of

1D line segment intersections, see

displayed here.

TABLE 1. EXECUTION AVERAGE TIMES OF ALGORITHM IN SECONDS

Computer Science & Information Technology (CS & IT)

ESULTS

implemented in MacPython 3.3.3. The test time results are obtained via Python

Time tests were performed on Apple Macintosh OS X processor (2.2 GHz intel

Core i7). The average time for no intersection, single point intersection (0D), line intersection

are measured in seconds. The program is executed

. The intersection times shown in Table 1 are neither optimized

include classification of intersections also. Times are

integrated algorithm works reliably on all triangle pairs

synthetic triangles that have every possible type of intersection. For

he times for no intersection are averaged over 100 runs with three samples

point intersection six sample pairs are averaged over 100 runs, then four samples are


and averaged. The composite average intersection time is computed. The test time

give three sample run output figures for examples of single point

intersection, a line intersection and an area intersection. The Matlab software is used to draw the

The user interface allows user to select any type of the possible triangle

intersections and it displays the corresponding triangle and the intersection

the sake of space consideration, one of the 0D (single point) intersections, see Fig. 1

, see Fig. 2; and one of 2D area intersections, see


Fig. 1. Single Point Intersection

obtained via Python

ntosh OS X processor (2.2 GHz intel

line intersection

100 times on

are neither optimized

are shown as

triangle pairs. The test

every possible type of intersection. For

runs with three samples. Similarly

sample pairs are averaged over 100 runs, then four samples are


time statistics are

examples of single point

is used to draw the

possible triangle-

triangle and the intersection. For

Fig. 1, one of the

, see Fig. 3, are



Fig. 2. Line Segment Intersection

Fig. 3. Area Intersection

5. APPLICATIONS OF TRIANGLE-TRIANGLE INTERSECTION

Here we describe two of the applications where triangle-triangle intersection is directly applicable

and is used extensively: qualitative spatial reasoning and geometric modeling. It is not limited to

these two applications, other applications include virtual reality, and computer vision.

5.1 Qualitative Spatial Reasoning

In Qualitative Spatial Reasoning, the spatial relations are determined by the 9-Intersection/4-

Intersection model[7, 8]. That is, for any pair of objects A and B, the interior-interior intersection

predicate, IntInt(A, B), has true or false value depending on whether the interior of A and the

interior of B intersect without regard to precise intersection. Similarly IntBnd(A, B) represents

the truth value for the intersection of the interior of A and the boundary of B, and BndBnd(A,B)

represents the predicate for the intersection of the boundaries of A and B. These four qualitative

spatial reasoning predicates are sufficient to define RCC8 spatial relations[7].


In the application VRCC-3D+[13

will need to intersect pairs of only triangles

algorithm uses axis aligned bounding boxes (AABB) to determine the closest triangles which

may possibly intersect[13]. Table

subsequently can be used to resolve the eight RCC8 relations

oriented towards the outside of the object

associated predicate is true. If the truth test fails, then other triangles need to be tested

of triangles results in a true value, then the result is false.

TABLE 2. CHARACTERIZATION OF INTERSECTION PREDICATES

This characterizes the intersection predicates

5.2 Geometric Modeling

In geometric modeling, the

CAD/CAM, the objects are represented with

Intersection between 3D surfaces

triangles. Briefly, a pair of surfaces is subdivided using axis aligned bounding boxes (AABB)

until the surfaces are reasonably planar and boundi

triangulated and the triangles are tested for cross

intersection segments are linked together to form curves of surface

curves may be open curves, closed curves, or even a combination of both [

presented here is extremely useful for geometric modeling where intersection between objects

occurs thousands of times. For geometric applications cross intersection is most

obtain the line segment of intersection.

comprehensive surface/surface intersection algorithm

6. CONCLUSION

Herein we presented a single algorithm for

complete framework for determining and characterizing

contrast to other single track algorithms

cross or coplanar intersection using only

based on logical tests on paramet

Cartesian coordinates. Thus our algorithm

computes the intersection as a single point

The algorithm provides more information than required by spatial reasoning systems

Consequently, we hope the additional infor

presented herein will be useful in other related applications


[13], the boundary of an object is already triangulated; that is, we

will need to intersect pairs of only triangles. To reduce the computational complexity, the


Table 2 is a characterization of the intersection predicates, which

subsequently can be used to resolve the eight RCC8 relations. Here we assume all normals are

oriented towards the outside of the object. Each characterization in Table 2 describes when the

If the truth test fails, then other triangles need to be tested

of triangles results in a true value, then the result is false.

CHARACTERIZATION OF INTERSECTION PREDICATES

intersection predicates that help in resolving the RCC8 relations.

surface-surface intersection problem occurs frequently

CAD/CAM, the objects are represented with surface boundaries using the ANSI (Brep) model.

surfaces amounts to detecting and computing intersection between 3D

Briefly, a pair of surfaces is subdivided using axis aligned bounding boxes (AABB)

until the surfaces are reasonably planar and bounding boxes intersect. Then the plane surfaces are

triangulated and the triangles are tested for cross-intersection and the intersection computed.

intersection segments are linked together to form curves of surface-surface intersection.

open curves, closed curves, or even a combination of both [4]. The algorithm


For geometric applications cross intersection is most often

line segment of intersection. Detailed analysis and implementation

/surface intersection algorithm may be found in [4].

a single algorithm for the complete design and a robust implementation of a

complete framework for determining and characterizing the triangle-triangle intersections

algorithms, this approach is a generic technique to detect any type

using only one algorithm. The classification of intersections is

on parametric coordinates rather than computational arithmetic tests

our algorithm not only detects the existence but also classifies

a single point, a line segment, or an area whichever the case may be

more information than required by spatial reasoning systems

e hope the additional information including classification of 3D intersection

presented herein will be useful in other related applications.

, the boundary of an object is already triangulated; that is, we

tational complexity, the


is a characterization of the intersection predicates, which

Here we assume all normals are

describes when the

If the truth test fails, then other triangles need to be tested. If no pair

8 relations.

frequently. In

ANSI (Brep) model.

intersection between 3D

Briefly, a pair of surfaces is subdivided using axis aligned bounding boxes (AABB)

ng boxes intersect. Then the plane surfaces are

intersection computed. The

tersection. The

The algorithm


often used to

and implementation of the most

implementation of a

triangle intersections. In

technique to detect any type

of intersections is

rather than computational arithmetic tests in

classifies and

whichever the case may be.

more information than required by spatial reasoning systems.

3D intersection


REFERENCES

[1] Max J. Egenhofer, R. Franzosa, Point-Set topological Relations, International Journal of Geographical

Information Systems 5(2), pp. 161 - 174, 1991.

[2] Oren Tropp, Ayellet Tal, IlanShimshoni, A fast triangle to triangle intersection test for collision

detection, Computer Animation and Virtual Worlds, Vol17 (50), pp.527 - 535, 2006.

[3] G. Caumon, Collon - Drouaillet P, Le Carlier de Veslud C, Viseur S, Sausse J (2009) Surface - based

3D modeling of geological structures. Math Geosci 41:927–945, 2009.

[4] E.G. Houghton, Emnett R.F., Factor J.D. and Sabharwal C.L., Implementation of A Divide and

Conquer Method For Surface Intersections; Computer Aided Geometric Design Vol.2, pp. 173 - 183,

1985.

[5] Ahmed H. Elsheikh, Mustafa Elsheikh, A reliable triangular mesh intersection algorithm and its

application in geological modeling, Engineering with Computers, pp.1 - 15, 2012.

[6] D. A. Randell, Z. Cui, and A.G. Cohn, A Spatial Logic Based on Regions and Connection., KR, 92,

pp. 165–176, 1992.

[7] C. Sabharwal and J. Leopold, “Reducing 9-Intersection to 4-Intersection for Identifying Relations in

Region Connection Calculus”, Proceedings of the 24th International Conference on Computer

Applications in Industry and Engineering (CAINE 2011), Honolulu, Hawaii, Nov. 16-18, 2011, pp.

118-123, 2011.

[8] Mo¨ller T. A fast triangle - triangle intersection test. Journal of Graphics Tools, 1997; 2(2): 25–30.

[9] Held M. ERIT a collection of efficient and reliable intersection tests. Journal of Graphics Tools

1997; 2(4): pp. 25–44, 1997.

[10] Badouel Didier, An Efficient Ray - Polygon Intersection, Graphics Gems (Andrew S. Glassner, ed.),

Academic Press, pp. 390 - 393, 1990.

[11] Guigue P, Devillers O. Fast and robust triangle - triangle overlap test using orientation predicates.

Journal of GraphicsTools 2003; 8 (1): pp. 25–42, 2003.

[12] Chaman Sabharwal, Jennifer Leopold, and Douglas McGeehan,:Triangle-Triangle Intersection

Determination and Classification to Support Qualitative Spatial Reasoning, Polibits, Research Journal

of Computer Science and Computer Engineering with Applications Issue 48 (July–December 2013),

pp. 13–22, 2013.

[13] N.Eloe, J. Leopold, C. Sabharwal, and Z. Yin, “Efficient Computation of Boundary Intersection and

Error Tolerance in VRCC-3D+ ”, Proceedings of the 18h International Conference on Distributed

Multimedia Systems (DMS’12), Miami, FL, Aug. 9 - 11, 2012, pp. 67 – 70, 2012.


INTENTIONAL BLANK



DESIGN AND IMPLEMENTATION OF

INTEL-SPONSORED REAL-TIME MULTI-

VIEW FACE DETECTION SYSTEM

Bohua Gan1, Vincent Chang

1, Guanying Wang

1, Xiuli Pan

1, Guan Wang

1,

Naihai Zou2, and Felming Feng

2

1University of Michigan—Shanghai Jiao Tong University Joint Institute,

Shanghai, China [email protected]

2Intel Asia-Pacific Research and Development Ltd, Shanghai, China

[email protected]

ABSTRACT

The paper introduces a case study of design and implementation of Intel-sponsored real-time

face detection system conducted in University of Michigan—Shanghai Jiao Tong University

Joint Institute (JI). This work is teamed up totally by 15 JI students and developed in three

phases during 2013 and 2014. The system design of face detection is based on Intel High

Definition (HD) 4000 graphics and OpenCL. With numerous techniques including the

accelerated pipeline over CPU and GPU, image decomposition, two-dimensional (2D) task

allocation, and the combination of Viola-Jones algorithm and continuously adaptive mean-shift

(Camshift) algorithm, the speed reaches 32 fps for real-time multi-view face detection. Plus, the

frontal view detection accuracy obtains 81% in Phase I and reaches 95% for multi-view

detection, in Phase III. Furthermore, an innovative application called face-detection game

controller (FDGC) is developed. At the time of this writing, the technology has been

implemented in wearable devices and mobile with Intel cores.

KEYWORDS

Multi-view face detection, real-time, graphic processing unit (GPU), OpenCL, face-detection

game controller (FDGC).

1. INTRODUCTION

Mobile commerce has reached unprecedented height: global smart phone shipments have over

1.16 billion in 2014. Meanwhile the competition of mobile processor chip market is fierce

between Qualcomm, MediaTek (MTK), Samsung, and Intel, etc. Qualcomm and MTK achieve

above 30 percent of market share each, following by Samsung’s 20 percent of the share. Unlike

its predominant advantage in the PC fields, Intel gains only 2.81 percent share in 2014 which

nearly missed the mobile revolution [1].


Intel’s core competitiveness hinges up advanced semiconductor fabrication process and

technology. In addition, effective and powerful software solutions make Intel the leading position

in the semiconductor industry. Furthermore, Intel has various System on Chip (SoC) for different

platforms and billions expenditures of Research & Development every year. Advanced

technologies such as SoC fabrication, Bluetooth, and wireless charging etc. are widely used in

Intel-inside wearable and mobile devices.

RealSense technology which incorporates face recognition and three-dimensional (3D) cameras is

used on the robots, unmanned aerial vehicles, as well as a new interactive mode of 3D immersive

experience. This innovative mode could eventually make user free from the constraint of

keyboards and passwords.

As the most important part of RealSense technology, face detection has been in-depth studied by

Intel. This corporation has sponsored three-phase real-time face detection capstone design

project (CDP) on University of Michigan－Shanghai Jiao-Tong University (Joint Institute)

including 15 students during 2013 and 2014 [2]. It aims at innovating in the face detection

technology and improving the performance of detecting speed, and accuracy etc.

The speed of original version which at start of Phase I is 3 fps for 480p videos frontal view face

detection and it reaches 9 fps at the end of Phase I. With the combination of image decomposition

and two-dimensional (2D) task allocation in Phase II, the speed is increased by 3 times and

reaching 30 fps. Combining the Camshift tracking in Phase III, multi-view face detection speed is

32 fps finally. Plus, the detection accuracy obtains 81% in Phase I and reaches 95%, in Phase III.

In addition, an interesting interactive application called Face-Detection Game Controller (FDGC)

is developed. The program’s conception, design details and testing results are presented in

Section 2. Conclusions are given in Section 3.

2. DESIGN AND IMPLEMENTATION OF THE REAL-TIME MULTI-VIEW

FACE DETECTION SYSTEM

As Table 1 shows, there are three phases in the face detection CDP, including two phases in 2013

and one phase in 2014. Every phase lasts about 12 weeks and involves 5 students.

2.1. OpenCL Accelerated Face Detection (Phase I, summer, 2013)

The first step is a literature survey and an objective analysis. According to the capability of

Intel’s HD 4000 graphics and OpenCL, the customer’s requirements are definite; they are

performance improvement, working prototype, and deliverable product separately. To meet the

customer requirements, the detailed engineering specifications are generated in Week 1. There

are mainly three aspects which are accuracy, efficiency, and innovative application [3]. As Table

2 shows, the objective of efficiency is reaching over 10fps for 480p videos, detecting 85% frontal

view faces correctly, and implementing an innovative application with bulged faces detection.

In Week 2, the entire video processing flows are definite as shown in Figure 1.The entire

processing function is decomposed into three workflows: video pre-processing, face detection

and accelerated process &innovative apps.


• Video Processing comprises of a series of preprocessing on the input video. It changes

the video stream into a different type of data which will make subsequent computation

easier.

• Face detection deals with the computation of images to locate human faces and find the

accurate sizes of them.

• Accelerated process will improve the speed exponentially and innovative applications are

concerned with how to implement application programs after face detection is done.

Through brainstorming, in which every member expressed their ideas of possible solution

concept related to the projects and Quality Function Deployment according to the discussion, the

combination of using Segment into Shots, Viola-Jones Algorithm, and GPU acceleration was

selected as the final design.

Table 1. Overview of all the tasks involved in developing this 3-phase work during 2013 and 2014. The

task marked with * represents the primary contribution for acceleration.

Week(nth) Phase I

(Summer, 2013) Phase II

(Fall, 2013) Phase III

(Summer, 2014)

1 Literature Survey, Objective Analysis

2

Function Decomposition

&Algorithm Choosing Previous Work Rebuilding

3 Image Face Detection

4 Video Stream Face

Detection Analysis & Design Skin-Color-Based Image

Segmentation Implementation

*

5

6 Performance Improvement

Pipeline of Accelerated

Concept over CPU and GPU

& Image Decomposition

and Task Allocation

*

7 Innovative Applications

Design

Viola-Jones detection + Camshift tracking Pipeline

Building* 8

9

Debug & Test Tuning Face Recognition using

Principal Components Analysis

* 10

11

Final Delivery

UI Design, User Feedback and

Update

Detection Box Merge Optimization & Image Preprocessing Pipeline

Tuning*

12 Final Delivery


Table 2. Engineering specifications of Phase I project, OpenCL acceleration face detection

Engineering Specifications Description

Accuracy Detect 85% faces correctly

Efficiency 10-15 fps for 480p videos with frontal view face detection

Innovative Application Adding bulged effect on the detected faces

Fig. 1. Video processing flow in Phase I

(a) (b)

Fig. 2. A special effect on the detected face in Phase I: (a) Original image (b) The image added with

the bulged effect on the detected face (Courtesy of Ref. [3]).

• Segment into shots, images are extracted on a certain rate and the number of extracted images

is reduced.

• Viola-Jones algorithm, it provides fast and efficient face detection framework [4].

• GPU acceleration, images separated from video streams can be dealt in parallel with GPU.

Therefore the speed is increased

From Week 3 to Week 5, the image and video stream face detection are completed. Due to slow

data exchange between CPU and GPU and the many sub-images that needed to be processed, the

performance of the initial version running in Intel HD 4000 graphics did not meet the

expectations. In order to improve the speed and accuracy, processing all sub-images in parallel,

exchanging data only when necessary, and buffering small exchanges are applied. Therefore, the

frontal view face detection speed has changed from 3 fps to 9 fps for 480p videos which is little

less than expectation. The accuracy is 81% in average.


Innovative application design is an important part of this project. And according to the mentor

from Intel, students are encouraged to think outside of the rules. The team implements bulge

effects on the detected faces as shown in Figure 2, and faces in the video are all bulged smoothly.

In Week 9 and 10, the team focused on testing and debugging. They found several bugs when

running with Intel hardware. This helps Intel improve the performance of the hardware, and it is

also beneficial to the following capstone design teams.

2.2. OpenCL Accelerated Face Detection (Phase II, fall, 2013) Intel requires the team of Phase II to bring the Phase I face detection program to a new level with

the help of Intel Beignet OpenCL implementation. Specifically, two specific requirements were

mentioned by Intel:

• Optimize the existing parallel computing algorithm.

• Come up with innovative ideas on applications and realize some of them [5].

Based on first three weeks’ literature survey, objective analysis, and previous work rebuilding,

the engineering specifications are definite as shown in Table III. The goal of this phase is faster,

more accurate and innovative.

The pipeline of accelerated concept over CPU and GPU has displayed in Figure 3 and was

decided particular by the team in Weeks 4 and 5. The general process is to first turn video

streams into still frame which will divert into gray images in CPU. Classifier will get the image

faces data from gray images, which include information of the size and location, and then add

special effects on the faces in GPU. Thus making GPU and CPU work simultaneously, the speed

increases obviously and reaches 20 fps for 480p videos.

In order to further accelerate the process, some ideas were considered and two of them were used.

They proved to be available in Phase I. In this phase, the extensive improved solutions, including

image decomposition and task allocation, are designed in detail. Image decomposition is to make

pre-processing and classifiers more efficient, and task allocation is to make the acceleration over

CPU and GPU more efficient [6]. Due to the effort in these two factors, it reaches 30fps for 480p

videos which is higher than Phase I and the accuracy achieves 94% in average.

(1) Image decomposition. Since a GPU has the property of parallel calculation, the images were

separated into small parts and then transformed together into the GPU. With the little parts

calculated at the same time, the calculated speed is raised remarkably.

(2) Task allocation. The tasks are sorted in a chain before, but the cores of GPUs are distributed

like a 2D matrix. It takes quite a lot of time to match up each task with each core of a GPU. In

this phase, tasks are put into a 2D task matrix where each task is matched up with the cores in a

GPU automatically, and the matching time is greatly reduced.

Innovative application is the vital engineering specification in Phase II, so the team pays much

more attention on user interface and creative application design. In the final version of the

program, it supports two types of inputs, which are still images and video streams. As shown

in Figure 4, users can click the “Video” button or “Image” button to choose different input

forms. For choosing video as input, users need to make sure that the computer has been

connected to a web camera firstly. The program will do real-time face detection on the videos

42 Computer

captured by the camera. For choosing image as input, users need

as the input document.

Table 3. Engineering Specifications in Phase II.

Engineering Specifications

Accuracy

Efficiency

Innovative Application

Fig. 3. Design of the pipeline for face detection acceleration over CPU and GPU in Phase II.

Fig. 4. User Interface of the accelerated face detection system in Phase II (


captured by the camera. For choosing image as input, users need to click the “Browse” button

Table 3. Engineering Specifications in Phase II.

Description

Detect 90%-95% faces correctly

20-30 fps for 480p videos with frontal view faces detection

Adding different effects on the detected faces


Fig. 4. User Interface of the accelerated face detection system in Phase II (Courtesy of Ref. [

to click the “Browse” button

with frontal view faces detection


Courtesy of Ref. [6]).


(a) (b) (c)

Fig. 5. Specialeffects on the detected faces in Phase II: (a) Rectangle (b) Mosaic (c) Comic relief (Courtesy

of Ref. [6]).

In addition, users can click three options for face displaying as shown in Figure 5. They are the

rectangle, mosaic and comic relief effects.

2.3 OpenCL based Face Detection and Recognition Algorithms Tuning (Phase III,

summer, 2014) Different from previous phases, Phase III is asked to realize the tilted face detection, and

detecting multi-view faces makes the product satisfactory with the commercial requirements as

shown in Table 4 [7].

Three image pre-processing steps which have shown in Figure 6 are used to remove background.

Detection area is significantly reduced, and the speed is boosted with these techniques.

• Skin Segmentation is used to identify skin region. Firstly, RGB image is converted to

YCrCb image which ignores illumination, then skin model is applied to obtain segmented

binary image. Meanwhile, some topological methods, like dilation, erosion and Gaussian

smooth, are used to denoise and smooth the binary image [8].

• Bounding Boxes Finding is used to find the located detection area. Label the connected

white components in the binary image at first and find contours of each component. Then

rectangle windows are drawn to bounding each contour.

• Merge the overlapped boxes into larger joint boxes. The merge operation can reduce

duplication caused by the overlap which will reduce the execution time.

Table 4. Engineering Specifications in Phase III.

Engineering Specifications Description

Accuracy Detect 90%-95% faces correctly

Efficiency Over 30 fps for 480p videos with multi-views detection

Functionality Tilted face detection

Innovative Application Face-Detection Game Controller

From Week 7 to Week 10, the team focused on increasing the accuracy, solving the problem of

detecting tilted faces and video face detection which based on the image face detection. Camshift

(Continuously Adaptive Mean-shift) [9] can detect tilted face in a large scale of angle and the

44 Computer

inner core of it involves more float and matrix calculation than Viola

builds a pipeline that leverages the consistency of video with Viola

algorithm. The new pipeline for the video

three steps as follows:

• Using Viola-Jones Algorithm to detect faces every 10 frame.

• Using Camshift Algorithm to track those faces for 9 continuous frames.

• Iterate previous two steps.

Skin

Segmentation

Input Images

Fig. 6. Flow chart for image pre

Fig 7. New pipeline combining Viola

Fig. 8. An innovative


inner core of it involves more float and matrix calculation than Viola-Jones. The solution which

builds a pipeline that leverages the consistency of video with Viola-Jones and Camshift

algorithm. The new pipeline for the video multi-view face detection shows in Fig. 7 and there are

Jones Algorithm to detect faces every 10 frame.

Using Camshift Algorithm to track those faces for 9 continuous frames.

Iterate previous two steps.

Bounding Boxes

Merging

Bounding Boxes

Finding

Output Images

Fig. 6. Flow chart for image pre-processing in Phase III (Courtesy of Ref. [7]).

Fig 7. New pipeline combining Viola-Jones Face Detection with Camshift Tracking (Courtesy of Ref. [

Fig. 8. An innovative application in Phase III: A face-detection game controller (FDGC)

(Courtesy of Ref. [7]).

Jones. The solution which

Jones and Camshift

view face detection shows in Fig. 7 and there are

Bounding Boxes

Merging

Output Images

Courtesy of Ref. [7]).

detection game controller (FDGC)


According this new pipeline, the speed breaks through to 32 fps for multi-views face detection

and the accuracy is over 95%. In order to show the improvement of Phase III and get real value

with new pipeline technic, the team applies it in Face-Detection Game Controller, in which

everyone can throw the keyboard, mouse away and play the video game by the face as shown in

Figure 8. Four more steps are needed for this extended pipeline:

• Detect & Track Faces, using pipeline to detect and track faces.

• Get the Angle of Faces, using geometrical principle to calculate the angle.

• Socket Communication between Windows & Unix, using socket communication to

control signals (i.e. Left/Right).

• Simulate Keyboard Signal, using windows inner function to simulate hardware signal

and operate the game.

3. CONCLUSIONS Intel sponsored 3-phase real-time face detection capstone design projects executed at JI during

2013—2014 in order to strengthen its RealSence technology. The ultimate goal of this work is to

catch up the market demands for mobiles and wearables as well as compete with predominant

industry giants. Prior to Phase I, the speed of the original version is 3 fps for 480p videos frontal

view face detection.

The platform development and system design of this work are based on Intel’s HD 4000 graphics

and OpenCL. It reaches 9 fps and 81% accuracy after the platform setup and system validation in

Phase I. With the techniques of image decomposition and 2D task allocation, the speed and

accuracy increase to 30 fps and 94%, respectively, for frontal view detection in Phase II.

Combining Viola-Jones algorithm with the Camshift tracking method, the speed reaches 32 fps

and the accuracy achieves 95% for real-time multi-view face detection in Phase III. Furthermore,

an interesting interactive application called face-detection game controller (FDGC) is developed.

The technology has been implemented in wearable devices and mobile with Intel cores at the

time of this writing.

ACKNOWLEDGEMENTS

The authors first thank Jackson He, Evelyn Yan, and Michael Fu from Intel for their financial and

technical support. Second, the authors thank Haoyang Chen, Guoxing Li, Zichao Fu, Xinyi Guo,

Sen Cao, Zhang Chen, Jifan Zhu, Xingmo Liu, Lili Su, Bochun Zhang, Shaoqing Zhu, and

Bowen Hong for their efforts in each phase project. Finally, the authors thank ChenyanFeng,

Taoxi Li, Jiayi Wu, and Shuo Zhang for the valuable discussion and assistance, Jason Chang for

editing the manuscript and Mian Li for helpful advice.

REFERENCES [1] R. James(2014), Brainstorm Tech conference in Aspen, Colorado. Available: http://fortune.com.

[2] V. Chang (2014), “MNC-Sponsored Multidisciplinary Industrial-Strength Capstone Design Projects

in China,” International Journal for Innovation Education and Research, Vol.2-10, pp 54-65.

[3] G. Wang, H. Chen, G. Li, Z. Fu, and X. Guo (2013), “OpenCL Accelerated Face Detection (Phase

I),” undergraduate thesis, University of Michigan—Shanghai JiaoTong University Joint Institute,

summer.


[4] P. Viola, and Michael J. Jones (2004), “Robust Real-Time Face Detection,” International Journal of

Computer Vision, Vol. 57, pp 137-154.

[5] X. Pan, S. Cao, Z. Chen, J. Zhu, and X. Liu (2013), “OpenCL Accelerated Face Detection (Phase II),”

undergraduate thesis, University of Michigan—Shanghai Jiao Tong University Joint Institute, fall.

[6] C. Fernandez, J. R. Saeta, X. Martorell, and J. Hernando (2011), “Real-time GPU-based Face

Detection in HD Video Sequences,” IEEE International Conference on Computer Vision Workshops,

November.

[7] G. Wang, L. Su, B. Zhang, S. Zhu, and B. Hong (2014), “OpenCL Based Face Detection and

Recognition Algorithm Tuning (Phase III),” undergraduate thesis, University of Michigan—Shanghai

Jiao Tong University Joint Institute, summer.

[8] H. Han, S. Shan, X. Chena, and W. Gao (2013), “A comparative study on illumination preprocessing

in face recognition,” Pattern Recognition, Vol. 46, pp 1691–1699.

[9] C. Du, H. Zhu, L. Luo, J. Liu, and X. Huang (2013), “Face detection in video based on AdaBoost

algorithm and skin model,” The Journal of China Universities of Posts and Telecommunications,

Volume 20, Supplement 1, pp 6–9, 24.

AUTHORS

Bohua Gan is a postgraduate student of Shanghai Jiao Tong University majors in

Electrical and Communication Engineering. He received B. E in Communication

Engineering from Dalian Maritime University in 2013 and LL. B in Law from Dalian

Maritime University in 2012. He is interested in all kinds of techniques, such as open

source GPGPU, antenna, and telemedicine etc.

He is also a research assistant of Prof. Vincent Chang since September, 2014. He

made effort in Intel-sponsored Capstone Design Projects focusing on OpenCL related

tasks on Intel GPU. He wishes to make our lives better through the technology.

Vincent Chang, Ph.D. & MBA is Teaching Professor & Faculty Director of

Corporate Relations at University of Michigan—Shanghai Jiao Tong University Joint

Institute (JI). He is also the founder of Knowledge Master, Inc. (KMI) in the US and

the cofounder of Semiconductor Leadership programs at UC Berkeley Extension.

Dr. Chang has been working in academia and industrial positions for more than 20

years in total with experiences in China, US and Taiwan. Dr. Chang pioneered

Corporate Partners Programs at JI, covering entrepreneurial innovation and industry

research. He strategically initiated and developed 96 MNC-sponsored Capstone

Design Projects focusing on healthcare, energy, electronics, and mobile internet,

creating an innovative platform collaborative with Covidien, GE, HP, Intel, Philips, and Siemens.

From 2004 to 2010, Dr. Chang was the President and CEO of KMI, a California-based e-learning company

specializing in semiconductor IC design and innovation leadership. The company was recognized as an

outstanding on-demand technology and content provider by UC Berkeley and awarded 9-year business

partner contract from 2007 to 2016. Prior to moving to the US in 2003, Dr. Chang was Associate Professor

of Electrical Engineering at Tamkang University in Taiwan. In 1995, he cofounded Calvin Engineering

specializing in structural design along with Calvin Chang in Taiwan in 1995. His research has appeared in

Electronics Letters, IEEE Journal of Lightwave Technology, IEEE Photonics Technology Letters, and

International Journal for Innovation Education and Research. He also authored 14 textbooks in

microelectronics published in Taiwan.


Guanying Wang is a software engineer at Google. She received B. E in Electrical

and Computer Engineering from Shanghai Jiao Tong University, Shanghai, China

and B.E in Computer Software Engineering from University of Michigan, Ann

Arbor, in 2013. She received M. E in Information Technology from Carnegie

Mellon University in 2014. She is interested in all kinds of novel technologies. She

had participated in Mobile Advertisement Syst

Application, Contextual Design, Inquires and User Interface Prototyping, Intel

Innovation Project: OpenCL Accelerated Face Detection and Carnegie Financial

Service Web App Development. And she is also a good manager of the p

Xiuli Pan is a student of Shanghai Jiao Tong University majors in Electrical and

Computer Science Engineering. He is interested in machines as well as programs

for the logic inside them is attractive.

He is interested in all kinds of techniques

automatic control and programming. With all of the skills I have learned I have

made a PRP system based on microprocessor and Bluetooth to detect the barrier and

guide the road on smart phones for blind people.

He always dream of making life better by making the techniques more easy to use

and more ways to be used. And this idea makes me to join this project and want to

make the face detection more easily to use and have a wild use in all areas.

Guan Wang was born in Beijing, China,

electrical and computer science engineering from Shanghai Jiao Tong University,

Shanghai, China and B.S.E. degree in computer science engineering from

University of Michigan, Ann Arbor, in 2014.

His research interests are are mainly in the deep learning, data mining, and

information retrieval. His is doing research study with Prof. Honglak Lee. Try to

build a system that detecting vehicle as well as road conditions by analyzing real

time video and pipelining multiple detection and tracking algorithms in machine

learning and computer vision, such as Convolution Neural Network (CNN), and Discriminative Part

Models (DPM). He is also a research assistant of Prof. Michael Cafarella. He made effort in Senbaz

which is a prototype spreadsheet database management system (SSDBMS), and also helped developing a

Feature Engineering Infrastructure that could dramatically ease the Explore

loop that characterizes many trained system pro

into deep learning and information retrieval area and leveraging those knowledge to make something

meaningful.

Nanhai Zou is a senior staff engineer in Intel Asia

center. He works for Intel Open source technology center. He is leading a team to

work on open source GPGPU and OpenCL related tasks on Intel GPU. Nanhai Zou

graduated from Shanghai Jiao Tong university. He holds a master degree of thermal

engineering.

Fleming Feng graduated from Tsinghua University at 1990, and then went to

Shanghai Jiao Tong University for a master’s degree. He started to work for Intel

from year 1997 after 4 year experience in a software company. From 2000, Fleming

started to work on Linux and open source project development. Ander from year

2002, Fleming began to work with TSP(the predecessor of OTC), managing the

development team for Carrier Grace Linux, and from 2005, took the role for

managing the OTC Shanghai team. From 2009, Fl

OSS scientist majorly focusing on Linux and open source collaboration with China

government, industry and university.


is a software engineer at Google. She received B. E in Electrical

and Computer Engineering from Shanghai Jiao Tong University, Shanghai, China

ftware Engineering from University of Michigan, Ann

Arbor, in 2013. She received M. E in Information Technology from Carnegie

Mellon University in 2014. She is interested in all kinds of novel technologies. She

had participated in Mobile Advertisement System Design, Dynamic Keyboard IOS

Application, Contextual Design, Inquires and User Interface Prototyping, Intel

Innovation Project: OpenCL Accelerated Face Detection and Carnegie Financial

Service Web App Development. And she is also a good manager of the project.

is a student of Shanghai Jiao Tong University majors in Electrical and

Computer Science Engineering. He is interested in machines as well as programs

for the logic inside them is attractive.

He is interested in all kinds of techniques such as network, microprocessor,

automatic control and programming. With all of the skills I have learned I have

made a PRP system based on microprocessor and Bluetooth to detect the barrier and

guide the road on smart phones for blind people.

eam of making life better by making the techniques more easy to use

and more ways to be used. And this idea makes me to join this project and want to

make the face detection more easily to use and have a wild use in all areas.

ing, China, in1992. He received B.S.E. degree in

electrical and computer science engineering from Shanghai Jiao Tong University,

Shanghai, China and B.S.E. degree in computer science engineering from

University of Michigan, Ann Arbor, in 2014.

research interests are are mainly in the deep learning, data mining, and

information retrieval. His is doing research study with Prof. Honglak Lee. Try to

build a system that detecting vehicle as well as road conditions by analyzing real-

elining multiple detection and tracking algorithms in machine


Models (DPM). He is also a research assistant of Prof. Michael Cafarella. He made effort in Senbaz


Feature Engineering Infrastructure that could dramatically ease the Explore-Extract-Evaluate interaction

loop that characterizes many trained system projects. As for his future plan, he wishes to continue digging


is a senior staff engineer in Intel Asia-Pacific research and developing

center. He works for Intel Open source technology center. He is leading a team to

work on open source GPGPU and OpenCL related tasks on Intel GPU. Nanhai Zou

graduated from Shanghai Jiao Tong university. He holds a master degree of thermal

graduated from Tsinghua University at 1990, and then went to

Shanghai Jiao Tong University for a master’s degree. He started to work for Intel

from year 1997 after 4 year experience in a software company. From 2000, Fleming

n Linux and open source project development. Ander from year

2002, Fleming began to work with TSP(the predecessor of OTC), managing the

development team for Carrier Grace Linux, and from 2005, took the role for

managing the OTC Shanghai team. From 2009, Fleming worked as the PRC chief

OSS scientist majorly focusing on Linux and open source collaboration with China

government, industry and university.

47


Models (DPM). He is also a research assistant of Prof. Michael Cafarella. He made effort in Senbazuru,


Evaluate interaction

jects. As for his future plan, he wishes to continue digging



INTENTIONAL BLANK



ENHANCING PERFORMANCE OF AN HPC

CLUSTER BY ADOPTING NON-

DEDICATED NODES

Pil Seong Park

Department of Computer Science, University of Suwon, Hwasung, Korea [email protected]

ABSTRACT

Persona-sized HPC clusters are widely used in many small labs, because they are cost-effective

and easy to build. Instead of adding costly new nodes to old clusters, we may try to make use of

some servers’ idle times by including them working independently on the same LAN, especially

during the night. However such extension across a firewall raises not only some security

problem with NFS but also a load balancing problem caused by heterogeneity. In this paper, we

propose a method to solve such problems using only old techniques applicable to old systems as

is, without requiring any upgrade for hardware or software. Some experimental results dealing

with heterogeneity and load balancing are presented using a two-queue overflow queuing

network problem.

KEYWORDS

HPC clusters, Security, NFS, SSH tunnelling, load balancing

1. INTRODUCTION

The desire to get more computing power and higher reliability by orchestrating a number of low

cost commercial off-the-shelf computers has given rise to a variety of architectures and

configurations. A computer cluster is composed of a set of loosely or tightly coupled computers

that are usually connected to each other through a fast LAN and work together so that they can be

viewed as a single system in many respects.

Computer clusters may be configured for different purposes. Load-balancing (LB) clusters are

configurations in which nodes share computational workload like a web server cluster. High-

performance computing (HPC) clusters are used for computation-intensive purposes, rather than

handling IO-oriented operations. High-availability (HA) clusters improve the availability of the

cluster, by having redundant nodes, which are then used to provide service when system

components fail.

Many kinds of commercial clusters are on the market. However the technologies to build similar

systems using off-the-shelf components are widely known (e.g., see [1]), and it is easy to build a


low-cost personal-sized cluster [2]. Many small labs first build their own personal-sized cluster,

and gradually grow it by adding some more dedicated nodes later. Instead, if there are some other

nodes that are being used for other purposes on the same LAN, we may try to utilize their idle

times as long as they are not always busy enough, especially during the night. The Berkeley NOW (Network of Workstations) project is one of early attempts to make use of the power of

clustered machines on a building-wide scale [3]. However such extension gives rise to difficulties

in security, administering the cluster, and load forecasting for optimal performance.

In this paper, we deal with some technical issues in extending a personal-sized Linux cluster

across a LAN. We do not use state-of-the-art technologies, since the sole purpose is to achieve

our purpose using our old HPC clusters as is, with no hardware or software upgrade. Some results

of the experiments to evaluate the system are given at the end.

2. BACKGROUND

2.1. HPC Cluster Middlewares

An HPC cluster normally consists of the dedicated nodes that reside on a secure isolated private

network behind a firewall. Users normally access the master/login node (master, for short) only,

which sits in front of compute nodes (slave nodes, or slaves for short).

The activities of all compute nodes are orchestrated by a cluster middleware that allows treating

the whole cluster system via a single system image concept. Well-known HPC middlewares

based on message passing are the Message Passing Interface (MPI) [4] and the Parallel Virtual

Machine (PVM) [5], the former being the de facto standard. According to the standard, many

commercial or non-commercial implementation libraries have been developed. LAM/MPI, FT-

MPI, and LA-MPI are some of widely used non-commercial libraries, and their technologies and

resources have been combined into the on-going Open MPI project [6].

All the nodes in an HPC cluster share not only executable codes but also data via NFS (Network

File System), which is not encrypted in general. It is perfectly all right as long as the whole

cluster nodes are on a secure local area network behind a firewall.

To extend the cluster to include other nodes outside of the firewall, we will be confronted with

some problems with security and data sharing among nodes. Moreover, such growth results in a

heterogeneous cluster with nodes of different power, possibly running different Linux versions.

2.2. File Sharing Protocols

NFS, created by Sun Microsystems in 1984, is a protocol to allow file sharing between UNIX

systems residing on a LAN. Linux NFS clients support three versions of the NFS protocol:

NFSv2 (1989), NFSv3 (1995), and NFSv4 (2000). However the NFS protocol as is has many

problems to use in extending the cluster, since its packets are not encrypted and due to other

shortcomings to be discussed later.

Other alternatives to NFS include AFS (Andrew File System), DFS (Distributed File System),

RFS (Remote File System), Netware, etc. [7]. There are also various clustered file systems shared


by multiple servers [8]. However we do not adopt such new technologies since they are not

supported by our old cluster.

2.3. SSH Tunnelling

Encryption of NFS traffic is necessary for secure extension across a firewall. One of the common

techniques that are ordinarily used is known as cryptographically protected tunnelling. An IP-

level or TCP-level stream of packets is used to tunnel application-layer segments [9]. A TCP

tunnel is a technology that aggregates and transfers packets between two hosts as a single TCP

connection. By using a TCP tunnel, several protocols can be transparently transmitted through a

firewall. Under certain conditions, it is known that the use of a TCP tunnel severely degrades the

end-to-end TCP performance, which is called TCP meltdown problem [10].

The SSH protocol allows any client and server programs to communicate securely over an

insecure network. Furthermore, it allows tunnelling (port forwarding) of any TCP connection on

top of SSH, so as to cryptographically protect any application that uses clear-text protocols.

3. EXTENSION OF AN HPC CLUSTER

We would like to extend our old cluster to include some other nodes across the firewall, as shown

in Figure 1. In general, the master node has attached storage that is accessible by diskless slave

nodes using insecure NFS. Since NFS relies on the inherently insecure unencrypted UDP

protocol (up to NFSv3), IP spoofing is possible. Moreover, firewall configuration is difficult

because of the way NFS daemons work.

Figure 1. Cluster extension to include the non-dedicated node EXT across a firewall.

3.1. Fixing NFS Ports for SSH Tunnelling

SSH tunnelling, which makes use of SSH port forwarding, is widely used to encrypt some

unencrypted packets or to bypass firewalls, e.g., see [9,11]. SSH tunnels support only TCP

protocols of fixed ports, but NFS uses UDP protocols by default, and the ports of some daemons

essential for the operation of NFS are variable.

Fortunately NFS over TCP protocols are also supported from the Linux kernel 2.4 and later on

the NFS client side, and from the kernel 2.4.19 on the server side [12]. Since all the nodes of our

cluster satisfy this condition, we can make NFS work over TCP by specifying the option "-o tcp"

in the mounting command. The following is an example of mounting server’s /nfs_dir directory

on the client’s mount_pt.


# mount -t nfs -o tcp server:/nfs_dir mount_pt

The daemons essential for NFS operation are 2 kinds. Portmapper (port 111) and rpc.nfsd (port

2049) use fixed ports, but rpc.statd, rpc.lockd, rpc.mountd, and rpc.rquotad use ports that are

randomly assigned by the operating system. However the ports of the latter can be fixed by

specifying port numbers in appropriate configuration files [13].

The ports of the first three can be fixed by adding the following lines in the NFS configuration

file /etc/sysconfig/nfs, for example,

STATD_PORT=4001

LOCKD_TCPPORT=4002

LOCKD_UDPPORT=4002

MOUNTD_PORT=4003

The rest ports can be fixed by defining new port numbers in /etc/services, for example

rquotad 4004/tcp

rquotad 4004/u에

3.2. Setting up an SSH Tunnel

For tunnelling of NFS, it is necessary for the server to mount its own NFS directory to be

exported to clients. Hence on the server side, the configuration file /etc/exports has to be

modified, for example, for exporting /nfs_dir,

/nfs_dir localhost (sync,rw,insecure,root_squash)

where “insecure” means it allows connection from ports higher than 1023, and “root_squash”

means it squashes the root permissions for the client and denies root access to access/create files

on the NFS server for security.

On the client side, to forward the ports 10000 and 20000, for example, to the fixed ports 2049

(rpc.nfsd) and 4003 (rpc.mountd), respectively, we can use the command

# ssh nfssvr -L 10000:localhost:2049 -L 20000:localhost:4003 -f sleep 600m

where "nfssvr" is the IP or the name of the NFS server registered in /etc/hosts, and "-f sleep

600m" means that port forwarding is to last for 600 minutes in the background.

Once connected to the NFS server, an SSH tunnel will be open if the correct password is entered.

Then the NFS server’s export directory can be mounted on the client’s side. The following

command is an example to mount the /nfs_dir directory of the NFS server on the /client_dir

directory of the client.

# mount -t nfs -o tcp,hard,intr,port=10000, mountport=20000 localhost:/client_dir


4. DEALING WITH HETEROGENEITY

Data partitioning and load balancing are important components in parallel computation. Since

earlier works (e.g., see [14]), many authors have studied load balancing using different strategies

on dedicated/non-dedicated heterogeneous systems [15,16,17,18], but it is nearly impossible to

find works on the security problems arising in cluster expansion, which is more technical rather

than academic.

Even though the original tightly-coupled HPC cluster may be homogeneous, the extended cluster

inevitably will behave like a heterogeneous system. First, the power of the newly added node

EXT may be different and its workload may change continually since it is not a dedicate node. In

addition, the communication speed between the node EXT and the original cluster is slower than

that among the original cluster nodes. Hence we need to use a dynamic run-time load balancing

strategy [19], while assigning equal amount of work to the nodes of the original cluster.

4.1. The Sample Problem

Strategies of dynamic load balancing depend on problems, and we need introduce the problem we

consider first. We deal with the old famous two-queue overflow queuing problem given by the

Kolmogorov balance equation

( ) ( )

( ) ( )1 2 2

1

1 , 1 , 1 2 , 1 1 1 2 2 ,

1 ,0 1 , 1 , 1 1 1 ,

[ 1 1 m in ( , ) m in ( , ) ]

= 1 1 m in ( 1, ) , , 1, 2 ,

i n j n j n i j

i i j i n i j

i s j s p

p i s p i j

λ δ δ λ δ µ µ

λ δ µ δ

− − −

− − +

− + − + +

− + − + =

where pi,j is the steady-state probability distribution giving the probability that there are ij

customers in the j-th queue. It is assumed that, in the i-th queue, there are si parallel servers and

ni-si-1 waiting spaces, and customers enter the i-th queue with mean arrival rate λi, and depart at

the mean rate µi. The model allows overflow from the first queue to the second queue if the first

queue is full. The total number of possible states is n1×n2.

It is known that no analytic solution exists, and the balance equation has to be solved explicitly.

We can write the discrete equation as a singular linear system, say 0A =x where x is the vector

consisting of all states pi,j’s. Even for systems with relatively small numbers of queues, waiting

spaces, and servers per queue, the size of the resulting matrix is huge. The numbers of waiting

spaces in our problem are 200 and 100 in each queue, respectively, and the resulting matrix size

is 20,000ⅹ20,000.

4.2. Job Assignment by Domain Decomposition

The graph of the resulting matrix of a k-queue problem is the same as the graph of the matrix

corresponding to the k-dimensional Laplacian operator. Thus the graph of the matrix A is a two-

dimensional rectangular grid.

As an example, Figure 1.a) shows the grid of a two-queue overflow model with 9ⅹ5 = 45 states,

where each dot corresponds to some state pi,j, i=1,2,…9 and j=1,2….5. If we use 3 computation

nodes, we may assign 15 grid points to each node. However to compute the values at boundary


points (in vertical dotted boxes), each node needs the values at the boundary points computed by

adjacent nodes. As a result, node 0 and node 2 should have enough memory to store 20 grid

points (or 4 vertical grid lines) and node 2 needs to store 25 grid points (or 5 vertical grid lines),

even though each node updates only 15 grid points (or 3 vertical grid lines). The extra grid points

are to store the values to be received from adjacent nodes.

Figure 1.b) shows partitioning of the corresponding matrix A of size 45ⅹ45, each submatrix Ai

being of size 5ⅹ45. For better load balancing, workload will be assigned in units of one vertical

line (that corresponds to the computation with one submatrix Ai). Since we deal with a huge

matrix problem of size 20,000ⅹ20,000, one unit workload is computation with 100ⅹ20,000

submatrix.

a) b)

Figure 2. Examples showing a) job assignment to 3 nodes, and b) matrix partitioning

for a 9ⅹ5 = 45 state two-queue overflow queuing model.

To solve the singular system 0A =x , we use a splitting method A=D-C where D is the diagonal

matrix that consists of all diagonal entries of A, and C=D-A. Then we can rewrite D C=x x

which leads to a convergent eigenvalue problem 1D C

− =x x . For convergence test, we use the

size of the residual defined by

1

2 2 2|| r || || ( ) || where || || 1.I D C−= − =x x

4.3. Communication for Asynchronous Iteration

In normal synchronous parallel computation, each node updates the values of the grid points

assigned to it in lockstep fashion, so that it can give the same solution as that of serial

computation. During computation, each node should communicate with its adjacent nodes to

exchange boundary point values, and all nodes should synchronize computation.

The workload of the added node EXT may change dynamically all the time, and it is

unpredictable and uncontrollable. If the node EXT is busy doing some other jobs, it cannot

communicate with its adjacent nodes in time, degrading overall performance by making all other

nodes wait.


Asynchronous algorithms appear naturally in parallel systems and are heavily exploited

applications.Allowing nodes to operate in an asynchronous manner simplifies the implementation

of parallel algorithms and eliminates the overhead associated with synchronization. Since 1990’s,

some theories on asynchronous iteration has been developed, e.g., see [20]. Fortunately, the

matrix A in our problem satisfies the necessary condition for a matrix to satisfy for convergence.

Hence we adopt asynchronous iteration freely, but we need some more work for implementation

details.

We assume the master node is always alive and is not busy so that it can immediately handle

communication with other nodes. Data exchange between two adjacent nodes should be done

through the master node, and the master node should keep the most current values of all grid

points. But all communication should be initiated by compute nodes (including the added non-

dedicate nodes), and the master node should just respond according to compute nodes’ request.

Figure 3 is the master’s algorithm using MPI functions to handle other nodes’ request. The

master continually checks if any message has arrived from any other nodes by calling the

MPI_Iprobe( ) function. If there is any message arrived, it then identifies the sender, tag, and

sometimes the number of data received. Based on the tag received, the master takes some

appropriate action. For example, if the received tag is 1000, it means “Receive the boundary data

I send, and send the boundary data I need.”, etc. Then any compute node can continue its work

without delay.

Figure 3. A master’s algorithm to handle compute nodes’ request.

We also adopt a dynamic load balancing strategy by measuring the time interval between two

consecutive data requests by the same node. It is necessary only for non-dedicated nodes, since

the others are dedicated and have the same computational power. The workload of the non-

dedicated node is reduced if

N tn

Tα<

where N is the average workload (in units of the number of vertical grid lines), T is the average

time interval between two consecutive data requests by the same dedicated node, t is the time

interval between two requests of EXT, and α is some tolerance parameter less than 1 (we used

0.8). Similarly, the workload of the non-dedicate node is increased if

N tn

Tβ>


for some β (we used 1.2).

In adjusting workloads, we used a simpler strategy. That is, when one of these conditions is

satisfied, we just decrease/increase two boundary grid lines of EXT and assigned them to/from

adjacent nodes. For this, to each non-dedicated node, we assigned the grid region which is

between the two regions assigned to dedicated nodes. It takes very little time to re-assign grid

lines to some other node, since our matrix is very sparse (just 5 nonzero entries in each row) and

can be generated by each node instantly whenever necessary.

5. PERFORMANCE TESTS

Our old cluster we want to extend has 2 nodes each with dual 1GHz Pentium 3 Xeon processors

running Fedora Core 4, with LAM/MPI v.7.1.2 installed. The nodes are interconnected via a

Gigabit LAN, and NFS is used for file sharing. The non-dedicated node EXT which sits on the

outside fast Ethernet has a 2.4 GHz Intel® Core™ i5 CPU with 8 GB memory, and runs on

Fedora 13.

The performance of the NFS through an SSH tunnel will inevitably drop due to encryption

overhead. Tests were performed between the master node and the node EXT, using UDP or TCP,

and with or without an SSH tunnel across the firewall. The times it took to read or write a file of

size 1GB from EXT were measured, at varying NFS block sizes. Since

NFSSVC_MAXBLKSIZE (maximum block size) of the NFS in Fedora Core 4 is 32*1024 (see

/usr/src/kernels/2.6.11-1.1369_FC4-i686/include/linux/nfsd/const.h), tests were performed at

block sizes of 4k, 8k, 16k, and 32k, respectively, 3 times for each and they are averaged. In

addition, to delete any remaining data in cache, NFS file system was manually unmounted and

mounted again between tests.

The following are example commands that first measure the time taken to create the file

/home/testfile of size 1GB on the NFS server and read it using block size 16KB.

# time dd if=/dev/zero of=/home/testfile bs=16k count=65536

# time dd if=/home/testfile of=/dev/null bs=16k

The results of the NFS performance test, with or without SSH tunnelling are given in Table 1. For

the common NFS without tunnelling, the figures in parentheses are the times it took when TCP is

used, and others are when UDP is used.

As we see, the larger the NFS block size, the faster in all cases. For the common NFS without

SSH tunnelling, write operation using UDP is slightly faster than TCP, but it is not the case for

read operation. Moreover the NFS with SSH tunnelling takes 3.9%-10.1% more time for write

and 1.4-4.3% more for read, than the common NFS using UDP.

As long as the NFS block size is taken as large as possible, the tunnelling overhead may not be

large even though NFS service is done through SSH tunnelling, since the non-dedicated nodes

outside of the firewall need not read or write so often through NFS, which is common in high

performance parallel computing.


Table 1. Performance of NFS, with or without SSH tunnelling (sec)

Block

size

Common NFS SSH tunnelled

Read Write Read Write

4K 96.12

(95.31)

119.37

(120.95) 100.29 131.45

8K 95.07

(94.43)

115.74

(120.31) 98.12 122.13

16K 93.85

(92.21)

114.70

(118.32) 95.31 121.72

32K 92.51

(91.45)

112.35

(115.87) 93.89 116.83

Figure 4 shows the comparison results. The solid line marked with 1 is the synchronous

computation result using only the dedicated nodes of our old cluster which has 4 cores in total.

The dotted line marked with 2 is the result when we add the node EXT (we created just 1 process

on EXT). The residual and elapsed time were checked every time the first slave reports the

intermediate result. The running times to converge to the result with residual size less than 10-5

were 956.75 sec for curve 1 and 854.14 sec for curve 2, respectively. Hence the speedup by the

added EXT node is 1.12.

Figure 4. Comparison results.

The result is somewhat disappointing, since the speed of the CPU core on EXT is at least twice as

fast as that of the old cluster. Probably it is mainly due to much slower network speed between

the cluster and the node EXT and the extra workload for some other jobs.

6. CONCLUSIONS

Personal-sized HPC clusters are widely used in many small labs, because they are easy to build,

cost-effective, and easy to grow. Instead of adding costly new nodes, we can extend clusters to

include some other servers on the same LAN, so that we can make use of their idle times.


However, unlike a tightly-coupled HPC cluster behind a firewall, the resulting system suffers a

security problem with NFS which is vital for HPC clusters.

Of course there are many new good solutions using recent technologies. However we do adopt

such strategy, because they require upgrade of hardwares and/or softwares including an operating

system. Instead we devise a solution using SSH tunnelling, which can be applied to the old

system as is. Probably this approach may be helpful to many of small home-made cluster

systems. We also devised a dynamic load balancing method which is based on domain

decomposition and asynchronous iteration, using a two-queue overflow queuing network problem

of matrix size 20,000ⅹ20,000. The speedup obtained by using one extra process on a much

faster non-dedicated node is somewhat disappointing, mainly because of the slow fast Ethernet

speed between the cluster and the extra node. We expect much higher speedup if the outer

network is upgraded.

ACKNOWLEDGEMENTS

This work was supported by the GRRC program of Gyeonggi province [GRRC SUWON2014-

B1, Cooperative Recognition and Response System based on Tension Sensing].

REFERENCES

[1] G. W. Burris, "Setting Up an HPC Cluster", ADMIN Magazine, 2015, http://www.admin-

magazine.com/HPC/Articles/real_world_hpc_setting_up_an_hpc_cluster

[2] M. Baker and R. Buyya, "Cluster Computing: the commodity supercomputer", Software-Practice and

Experience, vol.29(6), 1999, pp.551-576.

[3] Berkeley NOW Project, http://now.cs.berkeley.edu/

[4] M. Snir, S. Otto, S. Huss-Lederman, D. Walker, and J. J. Dongarra, MPI: The Complete Reference.

MIT Press, Cambridge, MA, USA, 1996.

[5] G. A. Geist, A. Beguelin, J. J. Dongarra, W. Jiang, R. Manchek, and V. S. Sunderam, PVM: Parallel

Virtual Machine: A Users' Guide and Tutorial for Networked Parallel Computing. MIT Press,

Cambridge, MA, USA 1994.

[6] Open MPI, http://www.open-mpi.org

[7] Linux NFS-HOWTO, http://nfs.sourceforge.net/nfs-howto/

[8] Clustered file system, http://en.wikipedia.org/wiki/Clustered_file_system

[9] M. Dusi, F. Gringoli, and L. Salgarelli, "A Preliminary Look at the Privacy of SSH Tunnels", in Proc.

17th International Conference on Computer Communications and Networks(ICCCN '08), 2008.

[10] O. Honda, H. Ohsaki, M. Imase, M. Ishizuka, and J. Murayama, "Understanding TCP over TCP:

Effects of TCP tunneling on end-to-end throughput and latency," in Proc. 2005 OpticsEast/ITCom,

Oct. 2005.

[11] Port Forwarding Using SSH Tunnel, http://www.fclose.com/b/linux/818/port-forwarding-using-ssh-

tunnel/

[12] Linux NFS Overview, FAQ and HOWTO, http://nfs.sourceforge.net/

[13] http://www.linuxquestions.org/questions/linux -security-4/firewall-blocking-nfs-even-though-ports-

are-open-294069/

[14] M. Zaki, W. Li, and S. Parthasarathy, "Customized Dynamic Load Balancing in a Heterogeneous

Network of Workstations", in 1996 Proc. 5th IEEE Int. Symposium on High Performance Distributed

Computing.

[15] M. Eggen, N. Franklin, and R. Eggen, "Load Balancing on a Non-dedicated Heterogeneous Network

of Workstations." in International Conference on Parallel and Distributed Processing Techniques and

Applications (PDPTA 2002), June 2002.


[16] J. Faik, L. G. Gervasio, J. E. Flaherty, J. Chang, J. D. Teresco, E.G. Boman, and K. D. Devine, "A

model for resource-aware load balancing on heterogeneous clusters", Tech. Rep. CS-03-03, Williams

College Dept. of Computer Science, http://www.cs.williams.edu/drum/, 2003.

[17] I. Galindo, F. Almeida, and J. M. Badia-Contelles, "Dynamic Load Balancing on Dedicated

Heterogeneous Systems", Recent Advances in Parallel Virtual Machine and Message Passing

Interface, Lecture Notes in Computer Science, vol.5205, 2008, pp 64-74

[18] K. Lu, R. Subrata, and A. Y. Zomaya, "On the performance-driven load distribution for

heterogeneous computational grids", J. of Computer and System Sciences vol.73, 2007, pp.1191-

1206.

[19] L. V. Kale, M. Bhandarkar, and R. Brunner, "Run-time Support for Adaptive Load Balancing", in

Lecture Notes in Computer Science, Proc. 4th Workshop on Runtime Systems for Parallel

Programming (RTSPP) Cancun - Mexico, vol.1800, 2000, pp.1152-1159.

[20] M. P. Raju and S. Khaitan, "Domain Decomposition Based High Performance Computing",

International J. of Computer Science Issues, vol.5, 2009,pp.27-32.

AUTHOR

Prof. Pil Seong Park received his B.S. degree in Physical Oceanography from Seoul

National University, Korea in 1977, and M.S. degree in Applied Mathematics from Old

Dominion University, U.S.A. in 1984. He received his Ph.D. degree in Interdisciplinary

Applied Mathematics (with emphasis on computer science) from University of Maryland

at College Park, U.S.A in 1991. He worked as the head of Computer Center at Korea

Ocean Research & Development Institute from 1991 to 1995. Currently, he is a professor

of the Department of Computer Science, University of Suwon in Korea. His research interest includes high

performance computing, Linux clusters, digital image processing, and knowledge-based information

systems. He is a member of several academic societies in Korea.


INTENTIONAL BLANK



ADAPTIVE AUTHENTICATION: A CASE STUDY

FOR UNIFIED AUTHENTICATION PLATFORM

Khairul Azmi Abu Bakar1 , Nor Izyani Daud

2

and Mohd Shafeq Md Hasan3

1Information Security Lab, MIMOS Berhad, Malaysia

[email protected] 2Information Security Lab, MIMOS Berhad, Malaysia

[email protected] 3Information Security Lab, MIMOS Berhad, Malaysia

[email protected]

ABSTRACT

Adaptive authentication is a risk-based authentication that identifies high-risk and suspicious

illegitimate login attempts. User past login records which implicitly contains attribute factors

context information are used to establish user behavior profile. Later if the user logins under

different environmental context from that established profile, the identity of the user may be

questioned. The system may challenge the user to present additional authentication method to

get authenticated. We implemented such adaptive authentication system in our production

server and collected user login records for more than six months. In this paper, we presents the

analysis of the user login profile with regards to attribute factors such as geographical location

and time of login. We also developed testbed system that uses the collected real data to evaluate

the system for different ratio threshold values.

KEYWORDS

Adaptive Authentication, Web Application, Testbed Analysis

1. INTRODUCTION

Authentication is a process to confirm that someone or something is, in fact, who or what it is

claimed to be. The process involves obtaining identification credentials or authentication method

such as username/password from a user and validating the credential against some authority.

Users who can present valid credential are considered authenticated identities. In general, there

are three categories of credentials: something you know (password), something you have (ATM

card) or what you are (fingerprint). To make it difficult for unauthorized person to gain access,

the system may implement multi-factor authentication where the user needs to successfully

present additional credentials from at least two of those three categories. If one factor is

compromised or broken, the attacker still has at least one more barrier to breach to break into the

system.


In traditional authentication system, the decision on the level of authentication credential required

solely depends on the application that the user trying to access. High sensitive applications such

as internet banking would demand the user to present stronger authentication credential than what

insensitive applications would. The required authentication methods could also be a combination

of two or more credentials, increasing the authentication security even more. However, in such

system, the security requirement is static because it only depends on the application security

requirement.

Adaptive authentication system uses login environmental characteristics and user behavioral

profiling to identify high-risk login and dynamically customizes the authentication requirement

accordingly. The system studies common behavior pattern of all users based on their past history

login access. If a user follows the same patterns when logging into the system, the login

experience may probably be a username/password indicating a low risk attempt. However, if a

user tries to login under different behavior or environment, the identity of the user is questioned.

The system may adaptively challenge the user to provide stronger or additional authentication

credentials to get authenticated.

We built such system in our lab and put it on our production server for couple of months. The

system had been collecting and storing users' login information into a database. We also

developed a testbed system that uses those data as the input parameters. The testbed enables us to

evaluate the behavior of our adaptive authentication system with different set of configurations

are used. In this paper, we present the result.

The remainder of the paper is organized as follows. Section 2 presents our current Unified

Authentication Platform (UAP) system that contains adaptive authentication component. Section

3 describes the formula to calculate User Attribute Score with regards to the relevant attribute

factors. Section 4 explains the algorithm used for the testbed environment that have been

developed to evaluate the system. Section 5 present the results from user login records analysis

and different ratio threshold values. Finally, Section 6 draws the conclusions.

2. BACKGROUND

In MIMOS Berhad, we have developed an authentication system called Unified Authentication

Platform (UAP). UAP is a centralized multi-factor authentication system with web-based single

sign-on (SSO) capability to manage user authentication profiles. It is designed to manage front-

end application authentication using an established protocol, Secure Assertion Markup Language

(SAML), which provides a centralized authentication framework and aims to reduce significant

application changes at the backend. The objectives of UAP are as the following:

1. provide an infrastructure that offers authentication service to applications

2. provide information technology that de-couples authentication function from application

and authorization

3. grow indigenous authentication mechanism industry throughout the country

4. a unified authentication platform initiative for enabling government e-services application

UAP is derived from Shibboleth [7] which is a standard based, open source package for web

single sign-on across or within organizational boundaries. In addition, UAP supports multiple

authentication methods. Users can choose from a list of authentication methods to get


authenticated and be allowed to access various applications without having to go through the

same authentication process again. The overall architecture of UAP is depicted in Figure 1[5].

Figure 1. Adaptive UAP Architecture Diagram

For the new generation of UAP, called adaptive UAP, we introduce an additional component

called Trust Engine that incorporates adaptive control based on security risk and level of

assurance. To make informed authentication decision, Trust Engine takes into account attribute

factors from the user behavior profiles which had been previously analyzed and stored.

Adaptive UAP consists of two processes: Pattern Generator and Trust Evaluator. Pattern

Generator analyses context information from user past login records which had been stored at

table data_log. Only login records from the last predefined period of time are processed. Context

information from the records such as login time and IP Addresses are converted into a

meaningful data format before the results along with the number of occurrence are stored at table

common_attr as the user behavior profile. Pattern Generator is currently configured to get

executed on every midnight.

Figure 2. Decision Factors


Trust Evaluator is responsible to analyze, decide and act upon every login request from users.

Trust Evaluator makes the final decision regarding on user authentication based on three factors

as shown in Figure 2: Authentication Method Score, User Attribute Score and Application

Required Trust Level. Authentication Method Score derives from the accumulated trust value of

all user successfully presented credentials. User Attribute Score is the result value from the

comparison process between the user current login contexts against the established user behavior

profile. Finally, Application Required Trust Level is the minimum trust level set by the

application that the user need to acquire to get authenticated. At the end of the process, Trust

Evaluator stores the login records into table data_log as depicted in Table 1 next to be processed

by component Pattern Generator.

Table 1. User Login Information.

No Entry Description

1. uuid User unique identity

2. time_login Data and time of the login

3. browser_osname User’s browser and OS information

4. ip_int User’s terminal IP Address

5. sp_id UAP Gateway Server ID string

6. auth_method Authentication method ID

7. Tr Application required trust level

Basically the requirement as in Formula 1 should be satisfied before the user is considered an

authenticated entity by the system.

� − � ≥ � (1)

Where

A: Authentication Method Strength

B: User Attribute Score

C: Application Security Requirement

If the established trust level does not meet the application required trust level, Trust Engine

returns FALSE to Authentication Server. User needs to present additional authentication method

to increase the established trust level. More explanation with some scenario examples can be read

at [3] and [4].

3. USER ATTRIBUTE SCORE

In this section, the second component of the formula that contributes for the final authentication

decision is discussed. User Attribute Score represents the uncertainty level of current login

attempt with respect to the established user behavior profile. In other words, if the user logins

under different behavior and environment from what the user normally had experienced, the score

value would be high. In this case study, attribute factors used are user geographical location (city

name), time login, type of browser and operating and targeted application.

There are two steps to calculate User Attribute Score. In the first step, the system needs to find

out the user common context for each attribute factor. Common context should meet the

following conditions:


1. the number of records for the last 14 days is more than 10

2. the frequency of occurrence of any particular context is more than a ratio threshold of

the overall records

Trust Evaluator should have gathered reasonably enough number of user records to establish

credible user attribute profile. In this case study, the minimum number of records required is set

to 10 from which only login records for the past 14 days are been considered .

To get qualified as a common context, the occurrence of the context should also exceeds a ratio

threshold. Presumably the threshold is set to 30%, if the user has 100 login records within the

period of time, any context that has at least 30 records is labeled as the user common context. For

example, if the user had been login from city of Kuala Lumpur for more than 30 times, Kuala

Lumpur is considered as the user common context under geolocation factor. If none of the

geolocation information meets both of the two conditions, the user is considered as not having

common context for that attribute factor and geolocation element is omitted from the formula.

Other attribute factors may still be applied subject to the same conditions.

If there exist common profile within the attribute factor, Trust Evaluator compares each of the

common profiles with the user current login context. If there is a match, the value of that attribute

factor is set to zero which gives no effect to the final score. Otherwise, the attribute factor is

activated. As explained later in this paper, the number of event where the attribute factor is

activated is recorded for analysis. The number reflects the rate of occurrence where the user

logins under different environment from the user behavior profile for every attribute factor.

Each of the attribute factor is assigned with a weightage value. The weightage value is in fraction

numbering format and represents the significance of the factor in user behavior profile. Higher

significant factor is assigned with a higher value. User attribute factor is calculated by adding the

weightage values for all activated factors as shown in (2) and multiply the result with a variable

max_user_score to limit the maximum possible scoring number.

attribute_score = (( time * weighttime)

+ (geolocation * weightlocation)

+ (browserOS * weightbrowser)

+ (application * weightapplication))

* (max_user_score) (2)

4. TESTBED SETUP

We have developed a testbed component for the Trust Engine that would able to take input

parameter from the stored records in table data_log instead of from the UAP server. Both

processes in Trust Engine take place as it is a real input data. The testbed allows us to evaluate

the performance of the Trust Engine when different set of configuration is used. Algorithm 1

shows the pseudo code for the testbed. Trust Evaluator component processes each of the login

record while Pattern Generator component is executed at the end of each day.


Algorithm 1. General Steps for TestBed

1: Generate list of days of table data_log

2: for Each day in the list do

3: for Every login records on day day do

4: process TrustEvaluator

5: end for

6: process PatternGenerator

7: end for

5. EXPERIMENT RESULT

We collected login information from our production server from 6 May 2014 until 15 January

2015. Total number of 171,045 login information from 1244 unique users have been recorded

during those 254 days period.

5.1. Attribute Factor

In this section, we present the overall analysis of the collected user login information based on

the attribute factor.

5.1.1. Geographical Location

The information about user geographical location is extracted from the IP Address of the user

terminal. We use solution from a third party company ip2location [6] which provides database

records that contain geographical information such as name of the city, region and country of

origin for IP Addresses. Special IP Addresses (10.0.0.0 to 10.255.255.255, 172.16.0.0 to

172.31.255.255, or 192.168.0 0 to 192.168.255.255) are originated from internal network and are

categorized as private IP Addresses.

From the total number of records collected, 67.7% (115,816) have IP Address information.

97.8% (113,254) of those login records that contains IP Address information are from internal

network. From the other 2,562 login records that comes from external network, 2,515 (98.1%) are

from Malaysia where city of Kuala Lumpur is the most originating login access location with

2,116 records (84.1%). The remaining login records are from USA (37), Sweden (3), Thailand

(3), Netherland (2) and Philippines (2).

5.1.2. Time Login

Time login entries stored in table data_log are the time of the server machine when the login

requests were received from UAP server. In other words, if the users login at a location that has a

different time zone, the server only stores the time of the server into the database. Time period is

divided into three time blocks, each with a different ID based on a standard working hours as

shown in Table 2.

The login records show that 95.4% of the login requests were received during the configured

working hours which is the second time block in the table. 2.9% of the total records are from the

first time block (7 pm- 12 am). Only 1.7% of the login access were recorded during the third time

block (12 am - 8 am). Figure 3 shows the number of login events based on the time hours.


Table 2. Time Block Duration

No Block ID Duration

1 A 12am – 8am

2 B 8am - 7pm

3 C 7pm – 12am

Figure 3. Number of Login with Respect to Time Hour.

The graph in Figure 3 also shows that the number of users login increases from 8 am onwards.

After 6:00 pm, the number of login starts to decrease. This is normal since the standard working

hours is between 8:30 am to 5:30 pm. In the afternoon, the number of login decreases for around

2 hours before increases back. It is a standard time for users to have their lunch break and stop

operation.

5.1.3. Type of Browser and Operating System

Adaptive UAP is able to extract information about the user terminal by looking at the user agent

string header send by the browser. There are ways to modify the string such as by using browser

extensions (User Agent Switcher [1] and User Agent Selector [2]). However, in this case study,

we assume that all user agent strings received by the system are original and unmodified.

Information such as type of browser and operating system running at the user terminal can be

determined based on the user agent string. Figure 4 shows an example of a user agent string. In

this example, the type of browser and operating system are Safari and iOS 8.1 respectively. It is

also clear that the user was using an iPhone smartphone.

Figure 4. Example of User Agent String Header.

Figure 5 and 6 show the distribution of type of browser and operating system the users used

during the data collection period. In term of type of browser usage as shown in Figure 5, majority

of login records are from browser Chrome (42.2%), followed by Firefox (33.7%) and Mozilla

(16.7%). From the operating system stand point, as shown in Figure 6, 92.2% of the login records

are from Windows operating system. The second most popular operating system is Mac OS

(4.3%), followed by Linux OS (2.2%).

Mozilla/5.0 (iPhone; CPU iPhone OS 8_1 like Mac OS X) AppleWebKit/600.1.4

(KHTML, like Gecko) CriOS/39.0.2171.50 Mobile/12B411 Safari/600.1.4


Figure 5. Number of Logins with Respect to Type of Browser

Trust Engine regards both entries of browser and operating system as one attribute factor which

carry one weightage value. If one of them is different, Trust Engine consider it as another entry.

Figure 7 shows the number of the paired value of browser and operating system derived from the

collected login records. The top four most favorite operating system is all Windows 7. In those

top list, the most popular browser is Chrome (36.4%), followed by Firefox (31.4%), Mozilla

(16.5%) and Internet Explorer (3.9%). The fifth place is browser Chrome with Windows 8

operating system which is 1.5% from the total number of collected login records.

5.1.4. Application

Six UAP gateway servers have been used with only one of them was intended for high trust

application. The threshold trust level for low trust and high trust application were set to 10 and 30

respectively. One UAP gateway server was purposely set to 0 so that the final authentication

decision by the Trust Engine is not affected by the attribute score value. Table 3 shows the list of

all UAP gateway servers used.

Figure 6. Number of Login with respect to Operating System


Figure 7. Number of Logins with Respect to paired Browser and OS

Table 3. List of all UAP Gateway Servers.

No UAP Gateway Server ID Required Trust

1. https://uap.mimos.my/shibboleth 0

2. https://sp.essweb.mimos.my/shibboleth 10

3. https://bmyhdw.mimos.com/8647196938 10

4. https://ipms-uap-gw-prod/vfdst35hj3120 10

5. https://misso.mimos.my/shibboleth 10

6. https://HT-MIESS1.mimos.my/shibboleth 30

There could be more than one application associated with one UAP gateway. If a user accesses

two different applications which are served by the same UAP gateway, Trust Engine gets the

same UAP gateway information for both login attempts.

Based on the collected login records, we found that majority (87%) of the logins were for

accessing UAP gateway labelled as https://sp.essweb.mimos.my/shibboleth. The second most

popular UAP gateway is https://HT-MIESS1.mimos.my/shibboleth which is assigned for high

trust application. High trust applications are used to view high confidential documents. Examples

of high trust applications are e-payslip to view the salary details online and e-PCB that shows

employer's monthly income tax payment to the government tax return agency.

5.2. Ratio Threshold

In this section, we present the effect of User Attribute Score when different configuration of ratio

thresholds are used. As explained in Section 3, the value assigned to the ratio threshold

determines the qualification of the common context for every attribute factor. The higher the

assigned value is, the more strict for any entry to be qualified as a common context. Each

individual user may have different pattern of behavioral profile.


The same collected data that had been stored in table data_log was used as the input parameters

to our testbed. The experiment setup allows us to see the effect of ratio threshold value to the

number of activated attribute factor for every login access. The configuration of the testbed is

shown in Table 4.

Table 4. Testbed Configuration

Entry Value

Attribute Weightage

1. Geographical location 8

2. Time 6

3. browserOS 4

4. Application 2

Credential Weightage

1. password 13

2. smsPin 20

3. otp 20

4. certificate 40

5. tck 20

6. tckbar 20

Time Interval 14 days

Threshold Ratio Percentage 10,30 and 50

The number of events when the attribute factors are activated is recorded and analyzed. In each of

those events, the respected attribute factors affect the outcome of the Attribute Score which then

reduces the final result of the user established trust. The summary of the results from the

experiment is depicted as in Figure 8. The last column labelled 'none' represents the login events

where no single attribute factor was activated. In this experiment we assume that there is no

adversary that wish to attack our system. In additional, all of the participating users were not

aware of the profile tracking and login under their own normal behavior.

We used different values (10, 30 and 50) for ratio threshold percentage to see their affect to each

Attribute Factor. From the Figure 8, we can see that in general for every attribute factor, the

higher the ratio threshold is, the higher number of events where the attribute factors are activated.

The number of events where no attribute factor is activated decreases when higher threshold ratio

percentage is used.

Type of browser contributes the highest number of events followed by application, operating

system, time login and location. We can conclude that users as individual have more tendency to

use different type of browser compared to other factors. Location is the least factor. One main

reason is because most of the users are the employers of the company and most of the

applications can only be accessed from the company private network.


Figure 8. Total Number of Events for every Attribute Factor.

6. CONCLUSION

Adaptive authentication is an additional security layer that uses risk factor analysis to make

authentication decision. Users' past login records are used to form user behavior profile. Context

information in login records such as user geographical location and time login are important

parameters to reflect user common behavior profile. If the user logins from a different

environment from the established behavior profile, adaptive authentication calculates the risk

associated with the deviation and may request the user to present additional authentication

method to be authenticated. We collects actual login records from the production environment for

more than 6 months. The login records profiling with regards to relevant attribute factors is

analyzed. We also develop a testbed environment that uses those records to evaluate our system

when different percentage of threshold ratios are used. In this paper, we present the results of the

experiment. For the future works, we plan to evaluate our testbed system for other parameters

setup.

ACKNOWLEDGEMENTS

We acknowledge the support provided by Ministry of Science, Technology and Innovation

(MOSTI) in funding the MIMOS Unified Authentication Platform (UAP) project through the

Tenth Malaysia Plan (10MP). The completion of the project allows the delivery of a centralized

authentication infrastructural platform for web applications.

REFERENCES [1] http://chrispederick.com/work/user-agent-switcher/. [Retrieved 30 January 2015].

[2] https://chromeuseragentselector.wordpress.com/. [Retrieved 30 January 2015].

[3] Khairul Azmi Abu Bakar and Galoh Rashidah Haron. Adaptive authentication: Issues and challenges.

In World Congress on Computer and Information Technology (WCCIT), pages 1–6, June 2013.

[4] Khairul Azmi Abu Bakar and Galoh Rashidah Haron. Adaptive authentication based on analysis of

user behavior. In Science and Information Conference (SAI), pages 601–606, August 2014.

[5] Galoh Rashidah Haron, Dharmadharshni Maniam, Vijayakumari Sadavisam, and Wong Hon Loon.

Re-engineering of web reverse proxy with shibboleth authentication. In The 7th International

Conference for Internet Technology and Secured Transactions (ICITST-2012), pages 325–330, 2012.


[6] ip2location. ip2location home page. http://www.ip2location.com. [Retrieved April 2012].

[7] Shibboleth. Shibboleth documentation. https://wiki.shibboleth.net. [Retrieved January 2014].

AUTHORS

Khairul Azmi Abu Bakar received the degree in Computer Engineering from Iowa

State University, USA in 1995 and master degree in Communication and Computer

from National University of Malaysia in 2002. He was awarded Ph.D. degree in

Electrical Engineering from University of Strathclyde, United Kingdom in 2012 for the

study on free-riding nodes in an open MANET. He is currently a staff researcher at

MIMOS Berhad where he has been since 1996. He has been involved in several R&D

projects in the field of micro-controller, smartcard, security systems under open source

platform. His primary research interests include wireless ad hoc security,

authentication system and computer network.

Nor Izyani Daud was born in Kuala Lumpur, Malaysia. She received the B.A Hons

Degree in Information Technology, Artificial Intelligence majoring from the Universiti

Utara Malaysia in 2000; and Master Degree in Real Time Software Engineering from

Universiti Teknologi Malaysia in 2006. In 2006, she joined MIMOS Berhad as a

Senior Engineer. She has been working in Information Security areas; for example

smart card programming, security scanning and analyzing. She also involved with

CMMI-Capability Maturity Model Integration implementation in the organization.

Mohd Shafeq Hasan was born in Kuala Lumpur, Malaysia in 1987. He received the

B.A Hons Degree in Information Technology, Java majoring from the Kuala Lumpur

Metropolitan University College in 2011, He joined MIMOS Berhad as a Junior

Engineer. He has been working in Information Security areas; for example adaptive

authentication, J2EE and Spring Security



SENSOR SELECTION SCHEME IN

WIRELESS SENSOR NETWORKS: A NEW

ROUTING APPROACH

Mohammad Alwadi1 and Girija Chetty

2

1Department of Information Sciences and Engineering,

The University of Canberra, Canberra, Australia [email protected]

2Faculty of ESTM, The University of Canberra, Canberra, Australia

[email protected]

ABSTRACT

In this paper, we propose a novel energy efficient environment monitoring scheme for wireless

sensor networks, based on data mining formulation. The proposed adapting routing scheme for

sensors for achieving energy efficiency. The experimental validation of the proposed approach

using publicly available Intel Berkeley lab Wireless Sensor Network dataset shows that it is

possible to achieve energy efficient environment monitoring for wireless sensor networks, with a

trade-off between accuracy and life time extension factor of sensors, using the proposed

approach.

KEYWORDS

Wireless sensor Networks, Physical environment Monitoring , machine learning, data mining,

feature selection, adaptive routing.

1. INTRODUCTION

Wireless and wired sensor networks (WSNs/SNs) has become a focus of intensive research

today, especially for monitoring and characterizing of large physical environments, and for

tracking environmental or physical conditions such as temperature, pressure, wind and humidity.

A wireless or a wired sensor network (WSN/SN) consists of a number of sensor nodes (few tens

to thousands) storing, processing and re-laying the sensed data, often to a base station for further

computation [1,2]. Sensor networks can be used in many applications, such as wildlife

monitoring [3], military target tracking and surveillance [4], hazardous environment exploration

[5], and natural disaster relief [6]. Many of these applications are expected to run unattended for

months or years. Sensor nodes are however constrained by limited resources, particularly in terms

of energy. Since communication is one order of magnitude more energy consuming than

processing, the design of data collection schemes that limit the amount of transmitted data is

therefore recognized as a central issue for wireless sensor networks. An efficient way to address

this challenge is to approximate, by means of mathematical models, the evolution of the


measurements taken by sensors over space and /or time. Indeed, whenever a mathematical model

may be used in place of the true measurements, significant gains in communications may be

obtained by only transmitting the parameters of the model instead of the set of real

measurements.

Since in most cases there is little or no a priori information about the variations taken by sensor

measurements, the models must be identified in an automated manner. This calls for the use of

machine learning and data mining techniques, which allow modelling the variations of future

measurements on the basis of past measurements.

In this paper, we introduce a novel data mining based formulation for energy efficient WSN

(Wireless Sensor Network) monitoring. The proposed approach involves an adaptive routing

scheme to be used for energy efficiency and is based on selecting most significant sensors for the

accurate modelling of the WSN environment. The experimental validation of the proposed

scheme for publicly available Intel Berkeley lab Wireless Sensor Network dataset shows it is

indeed possible to achieve energy efficiency without degradation in accurate characterization and

understanding of WSN environment. The proposed machine learning and data mining

formulation for achieving energy efficiency, provides better implementation mechanism in terms

of trade-off between accuracy and energy efficiency, due to an optimal combination of feature

selection and classifier techniques used in machine learning chain. By approaching the

complexity of WSN/SN with a data mining formulation, where each sensor node is equivalent to

an attribute or a feature of a data set, and all the sensor nodes together forming the WSN/SN set

up - equivalent to a multiple features or attributes of the data set, it is possible to use powerful

feature selection, dimensionality reduction and learning classifier algorithms from machine

learning/data mining field, and come up with an energy efficient environment monitoring system

[7]. In other words, by employing a good feature selection algorithm along with a good

classification algorithm, for example, it is possible to obtain an energy efficient solution with

acceptable characterization or classification accuracy (where the WSN/SN set up is emulated

with a data set acquired from the physical environment). Here, minimizing the number of sensors

for energy efficiency is very similar to minimizing the number of features with an optimal feature

selection scheme for the data mining problem. Further, the number of sensors chosen by the

feature selection scheme, leads to a routing scheme for collecting the data from sensors,

transmitting them until it reaches the base station node. As accuracy of data mining schemes rely

on amount of previous data available for predicting the future state of the environment, it is

possible to obtain an adaptive routing scheme, through the life of WSN, as more and more

historical data becomes available, allowing trade-off between energy efficiency and prediction

accuracy. We have shown that it is possible to do this, with an experimental validation of our

proposed scheme with a publicly available WSN dataset acquired from real physical

environment, the Intel Berkeley Lab [8]. Rest of the paper is organized as follows. Next Section

describes the details of the dataset used, and Section 3 de-scribes the proposed machine learning -

data mining approach. The details of experimental results obtained are presented in Section 4, and

the paper concludes in Section 5, with conclusions and plan for further research.

2. DATA SET DESCRIPTION

The publicly available data set used for experimental validation consists of temperature and

humidity measurements, which come from a deployment of 54 sensors in the Intel research

laboratory at Berkeley [8]. The deployment took place between February 28th and April 5th,


2004. A picture of the deployment is provided in Figure 1 below, where sensor nodes are

identified by numbers ranging from 1 to 54.

Figure 1. Intel Berkeley Wireless Sensor Network Data set: location of 54 sensors in an area of 1200m

2

Many sensor readings from WSN test bed were missing, due to this being a simple prototype. We

selected from this data set few subsets of measurements. The readings were originally sampled

every thirty-one seconds. A pre-processing stage where data was partitioned was applied to the

data set. After pre-processing, we prepared several subsets of data. The approach we used for this

WSN test bed, involves, an assumption that data collection and transmission is done by some of

the sensors (source nodes), that purely sense the environment and transmit their measurement to

collector nodes (sink nodes/base station), and based on the relative distance between source

nodes and sink nodes, the route or the path taken for sensor data to be transmitted from one node

to other is predetermined at sink/base station node. Those nodes who actively participate in

sensing the environment, and transmit the data, consume the power and those who do not

participate in this activity do not consume any power. This is how the WSN can be made energy

efficient; by involving optimum number of sensors to participate in environment sensing and

transmission task, and leaving non-participating sensors in sleep mode (no energy consumption).

This can however, impact on the accuracy of sensing the environment, if number of sensors

participating in routing scheme is not properly chosen. To ensure a trade-off between accuracy

and energy efficiency is achieved, it is essential that a dynamic or adaptive routing scheme is

used, where, the machine learning/data mining technique can use larger training data from

previous/historical data sets to predict the future environment accurately, and continuously adapt

the routing scheme for nodes based on a threshold error measure for prediction accuracy and

energy efficiency. Next Section describes the proposed machine learning data mining scheme

used for sensor selection and routing in this approach.

3. SENSOR SELECTION AND ROUTING APPROACH

The sensor selection and routing approach is based on a feature selection technique that selects

the attributes (sensors), by evaluating the worth of a subset of attributes by considering the

individual predictive ability of each feature/sensor along with the degree of redundancy between

them. Subsets of features that are highly correlated with the class while having low inter

correlation are preferred [9, 10]. Further, this feature/sensor selection algorithm identifies locally

predictive attributes, and iteratively adds the attributes with the highest correlation with the class


as long as there is not already an attribute in the subset that has a higher correlation with the

attribute in question. Once the appropriate group of sensors are selected, the prediction of sensor

output at sink node or base station is done by linear regression algorithm, using the Akaike

criterion [10, 11], which involves stepping through the attributes, removing the one with smallest

standardised coefficient until no improvement is observed in the estimate of the error given by

Akaike information metric. Figure 2 and the table below shows how the sensor selection evolves

as the training data (historic data) used for predicting the sink sensor output is increased, and

ensures the prediction accuracy/error is maintained at a particular threshold value.

Figure 2. Sensor selection map for Humidity Experiment 1,2 and 3

4. EXPERIMENTAL RESULTS AND DISCUSSION

Different sets of experiments were performed to examine the relative performance of sensor

selection and routing approach proposed here. We used k-fold stratified cross validation

technique for performing experiments, with k=2, 5 and 10, based on the training data available

(using larger folds for larger training data). Further, to estimate the relative energy efficiency

achieved, we performed experiments with all sensors (without feature selection/sensor selection)

algorithm, and with sensors selected by feature selection algorithm. As mentioned before, the

feature selection algorithm allows selection of an optimal number of features or sensor nodes

needed to characterize or to classify the environment (which in turn leads to an energy efficient

scheme). Further, time taken to build the model is also an important parameter, particularly for

adaptive sensor routine scheme to be used for real time environment monitoring.

Table 1. Results for Three experiments scenarios

Exp

number

Number of

Sensors

Number of

Training samples Features/sensors selected

1 52 35 7,24,41,44,50

2 52 2700 3,7,11,14,16,22,28,29,34,41

3 52 5400 14,19,24,25,36

Exp

number

Time (No F

selection) Time(F Selection)

RMSE No F

selection

RMSE with F

selection

1 0.01 0.01 3.82 0.04

2 0.14 0.01 0.96 2.11

3 0.17 0.03 1.91 4.56


For the first set of experiments, we used first 52 sensors and a small set of training samples (35

Humidity measurements). As can be seen from the sensor locations shown in Figure 2, sensor 50

is the sink node (emulating base station node), and sensors 1 to 49 participate in measuring and

transmitting the environment around them to the sink node, where the machine learning

prediction task is to estimate the measurement at sink node (sensor 50). The RMS error (root

mean squared error) at the sink node (node 50) provides a measure of prediction For all source

sensor nodes (1-49) in WSN participating in measuring the temperature in the environment and

sending it to sink node, the RMS error is 3.82%, and with sensor selection scheme used with only

5 sensors participating in routing scheme, the RMS error is 0.04%. As can be seen in Table 1,

with a moderate degradation in accuracy (3.82% to 0.04%), energy efficiency achieved is of the

order of 52 (52/5). We used a new measure for energy efficiency, the life time extension factor

(LTEF), which can be defined as:

LTEF =Total number of sensors

Sensors participating in the routing scheme

With 5 sensors out of 52 sensor nodes in active mode, the LTEF achieved is around 10 times, and

52 sensor nodes are in sleep mode. The trade off is a slight reduction in accuracy. This could be

due to less training data used. We used only 35 temperature samples for prediction scheme. With

more data samples used in the prediction scheme, performance could be better. To test this

hypothesis, we performed next set of experiments.

For second set of experiments, we used 2700 training samples collected on different days. As can

be seen in Table 1, with larger training data size, we found that the participating sensors in the

routing scheme are different, as the proposed feature selection algorithm chooses different set of

sensors (3,7,11,14,16,22,28,29,34,41). We used 52 sensors for this set of experiments, as two of

the sensors (sensor 5 did not have more than 35 measurements). With all 52 sensors in the routing

scheme, the RMS errors is 0.96%, and with 10 sensor nodes (3,7,11,14,16,22,28,29,34,41), the

error is 2.11%. This was not a significant improvement in prediction accuracy (from 0.96% to

2.11%), with life time extension of 5.2 (52/10). As is evident here, by using larger training data

(2700 temperature measurements), it was possible to achieve an improvement in prediction

accuracy and energy efficiency as well.

To examine the influence of increasing training data size, we performed third set of experiments

with 5200 samples. The performance achieved for this set of experiments is shown in Table 1.

Here the adaptive routing scheme based on proposed feature selection technique selects 5 sensors

(14,19,24,25,36). For this set of experiments, the RMS error varies from 1.91% for all sensors

participating in the scheme to 4.56 % with LTEF of 6.6 (52/5). Though there is no degradation in

prediction accuracy, there is not much improvement in energy efficiency, with doubling of

training data size for the building the model This could be due to overtraining that has happened,

with the network losing its generalization ability. So by increasing training data size, it may not

be just possible to achieve performance improvement, for pre-diction accuracy (RMS error) and

energy efficiency (LTEF), and a trade off may be needed. An optimal combination of training

data size, and number of sensors actively participating in routing scheme can result in energy

efficient WSN, without compromising the prediction accuracy.

Further, another important parameter is model building time, as for adaptive sensor routing

scheme to be implemented in real time WSN environment, routing scheme has to dynamically


compute the sensors that are in active mode and in sleep mode. Out of 3 experimental scenarios

considered here, as can be seen from Table 1, the model building time from 0.01seconds to 0.01

seconds for experiment 1, from 0.14 seconds to 0.01 seconds for experiment 2, and from 0.17

seconds to 0.03 seconds for experiment 3. So, the proposed adaptive routing scheme for sensor

selection provides an added benefit of reduced model building times, suitable for real time

deployment. figure below shows the time taken to build the model.

Figure 3. Time taken to build the model (Humidity Exp1,2 and 3)

5. CONCLUSIONS

Energy sources are very limited in sensor networks, in particular wireless sensor net-works. For

monitoring large physical environments using SNs and WSNs, it is important that appropriate

intelligent monitoring protocols and adaptive routing schemes are used to achieve energy

efficiency and increase in lifetime of sensor nodes, with-out compromising the accuracy of

characterizing the WSN environment. In this paper, we proposed an adaptive routing scheme for

sensor nodes in WSN, based on machine learning data mining formulation with a feature

selection algorithm that selects few most significant sensors to be active at a time, and adapts

them continuously as time evolves. The experimental validation for a real world publicly

available WSN dataset, proves our hypothesis, and allows energy efficiency to be achieved

without compromising the prediction accuracy, with an added benefit in terms of reduced model

building times. Further work involves, developing new algorithms for sensor selection and

environment characterization with WSNs and their experimental validation with other similar

datasets, that can lead to better energy efficiency. Also, our further re-search involves extending

this work with adapting these classifiers for big data stream data mining schemes, for real time

dynamic monitoring of complex and large physical environments in an energy efficient manner.

REFERENCES

[1] Ping, S., Delay measurement time synchronization for wireless sensor networks. Intel Research

Berkeley Lab, 2003.

[2] Hall, M., et al., The WEKA data mining software: an update. ACM SIGKDD Explorations

Newsletter, 2009. 11(1): p. 10-18.


[3] Csirik, J., P. Bertholet, and H. Bunke. Pattern recognition in wireless sensor networks in presence of

sensor failures. 2011.

[4] Nakamura, E.F. and A.A.F. Loureiro, Information fusion in wireless sensor networks, in Proceedings

of the 2008 ACM SIGMOD international conference on Management of data2008, ACM: Vancouver,

Canada. p. 1365-1372.

[5] Bashyal, S. and G.K. Venayagamoorthy. Collaborative routing algorithm for wireless sensor network

longevity. 2007. IEEE.

[6] Richter, R., Distributed Pattern Recognition in Wireless Sensor Networks.

[7] Alwadi, M. and G. Chetty, Energy Efficient Data Mining Scheme for Big Data Biodiversity

Environment. 2014.

[8] Bodik, P., et al., Intel lab data. Online dataset, 2004.

[9] Hall, M.A., Correlation-based feature selection for machine learning, 1999, The University of

Waikato.

[10] Akaike, H., Information theory and an extension of the maximum likelihood principle, in Selected

Papers of Hirotugu Akaike1998, Springer. p. 199-213.

[11] Ashraf, M., et al., A New Approach for Constructing Missing Features Values. International Journal

of Intelligent Information Processing, 2012. 3(1).

AUTHORS

Mohammad Alwadi

Mohammad Alwadi has bachelors degree in Computer Science from Jordan, Masters

degree from the University of Canberra, Australia and Currently a PhD candidate in

the Faculty of Information Science and Engineering at The University of Canberra,

Australia. Mohammad research interests are in the area of computer networking,

Data Mining and wireless sensor networks This author has 8 publications and

working towards the submission of his PhD Thesis at the University of Canberra.

Australia. He has more than 6 years of experience in the research field of Data

mining , sensors networks and wireless sensor networks.

Dr.Girija Chetty

Girija has a Bachelors and Masters degree in Electrical Engineering and Computer

Science from India, and PhD in Information Sciences and Engineering from

Australia. She has more than 25 years of experience in Industry, Research and

Teaching from Universities and Research and Development Organisations from

India and Australia, and has held several leadership positions including Head of

Software Engineering and Computer Science, and Course Director for Master of

Computing (Mainframe) Course. Currently, she is an Associate Professor, and Head

of the Multimodal Systems and Information Fusion Group in University of Canberra, Australia, and leads a

research group with several PhD students, Post Docs, research assistants and regular International and

National visiting researchers. She is a Senior Member of IEEE, USA, and senior member of Australian

Computer Society, and her research interests are in the area of multimodal systems, computer vision,

pattern recognition, data mining, and medical image computing. She has published extensively with more

than 120 fully refereed publications in several invited book chapters, edited books, high quality conference

and journals, and she is in the editorial boards, technical review committees and regular reviewer for

several IEEE, Elsevier and IET journals in the area related to her research interests. She is highly interested

in seeking wide and interdisciplinary collaborations, research scholars and visitors in her research group.


INTENTIONAL BLANK



AN GROUP BEHAVIOR MOBILITY MODEL

FOR OPPORTUNISTIC NETWORKS

GuoDong KANG1 and GuoLiang KANG

2

1DFH Satellite Co., Ltd., 100094, Beijing, China

[email protected] 2University of Technology, Sydney 15 Broadway, Ultimo NSW 2007

[email protected]

ABSTRACT

Mobility is regarded as a network transport mechanism for distributing data in many networks.

However, many mobility models ignore the fact that peer nodes often carried by people and

thus move in group pattern according to some kind of social relation. In this paper, we propose

one mobility model based on group behavior character which derives from real movement

scenario in daily life. This paper also gives the character analysis of this mobility model and

compares with the classic Random Waypoint Mobility model.

KEYWORDS

Mobility model , Group behavior , Opportunistic network

1. INTRODUCTION

Nowadays mobility is playing an important role more and more in many wireless networks such

as mobile ad-hoc networking (MANET) and delay-tolerant networking (DTN) [2, 3]. One

common important fact for these opportunistic networks is that mobile devices among them are

often carried by people. From this point of view, the mobility of mobile devices is exactly the

same as the movement of persons who are taking them. The idea of mobility structure in this

paper comes from people’s movement custom or character in everyday life.

Usually we can observe one familiar scenario: persons in the same group often go to some place

together then to another place together meantime everyone has his own walking or moving style.

How to construct this kind of mobility is the main work of this paper.

To construct this kind of mobility model, we have to face two questions: how to allocate the

given nodes into different groups according to certain method; how to describe and construct

nodes’ common group mobility behavior and every node’s individual mobility behavior. The

mobility model proposed in this paper thoroughly considers nodes’ individual realistic movement

character combing common group character based group division method.


The rest of this paper is organized as follows: Section II describes the allocation method of

mobile nodes in opportunistic network. Section III proposes the setup method of the mobility

models based on common group-behavior character and individual self-behavior character.

Section IV gives analysis of the proposed new mobility model. Conclusion is given in Section V.

2. GROUP ALLOCATION METHOD

The group allocation method can be undertaken as following two steps: Firstly, the network

should find and construct the relations among whole nodes in the form of mathematic matrix

which will be regarded as the basis of second step. Then whole nodes are allocated into different

groups according to the relations founded in first step.

2.1. Node Relation Setup

Normally, there is some kind of social or biological relations among those nodes since they are

carried by people. To set up this kind of mutual relation, we adopt the classical method of

representing social or biological network, weighted graphs. The strength of mutual relation

between any node pair is represented using a value in the range [0, 1]. As a consequence, the

network internal relation can be described as a relation matrix with a dimension of N*N where N

is the total number of nodes in the network. At present, there are several models that describe the

key properties of real-life network, such as random graph, small world, etc. Some research work

show that the properties of these random graphs, such as path lengths, clustering coefficients,

can’t be regarded as accurate models of realistic networks [5, 6, 7]. Here, we choose the

geometric random graph to be the network model. In this kind of model, the geometry relations

of nodes have strong association with the social relation of nodes. That means when any two

nodes are in the radio range of each other, the social relation exists. On the contrary, since they

even can’t communicate with each other, we think that there is no any social interaction between

them. So when the Euclidean distance between any two nodes is smaller than radio range R, the

corresponding element of the social relation matrix M is set to be 1 or else set to be 0 as follows.

,

1 ,

0

i j

i j

if i j and P P Rm

otherwise

≠ − ≤=

It should be emphasized that the relations value of one node and itself is regarded to be zero in

the matrix. In [6] it is shown that in two or three dimensional area using Euclidean norm can

supply surprisingly accurate reproductions of many features of real biological networks.

For example, 100 nodes are uniformly distributed as shown figure 1. Lines are used to represent

the mutual relations among these nodes whose radio range is assumed to be 10 meters here.

2.2. Nodes Group Allocation

Once the relation matrix M is obtained, groups can be detected. Group structure is one of the

common characters in many real networks. However, finding group structures within an arbitrary

network is an acknowledged difficult task. A lot of work has been done on that. Currently, there

are several methods that can achieve that goal, such as Minimum-cut method, Hierarchical

clustering, Girvan-Newman algorithm, etc.


Minimum-cut method is one of the oldest algorithms for dividing networks into parts. This

method uses in load balancing for parallel computing in order to minimize communication

between processor nodes. However, this method always finds communities regardless of whether

they are implicit in the structure, and it can only find a fixed number of them. So it is less than

ideal for finding community structure in general networks [4].

Hierarchical clustering is another method for finding community structures in network. This

method detects the community by defining a similarity measure quantifying some (usually

topological) type of similarity between node pairs.

Figure 1 An example of nodes relations. Small circles represent nodes. Short lines represent the relation

among them.

The Girvan–Newman algorithm is one commonly used algorithm for finding communities [12]. It

identifies edges in one network that lie between communities and then removes them, leaving

behind just the communities themselves. But this method runs slowly which makes it impractical

for networks of more than a few thousand nodes [11].

Modularity maximization is one of the most widely used methods for community detection [11].


into communities. This method detects the community structure of high modularity value by

exhaustively searching over all possible divisions [8].

In this paper we adopt modularity maximization as the social group detection method. Modularity

maximization is one of the most widely used methods for group detection [11].


into groups. This method detects the group structure of high modularity value by exhaustively

searching over all possible divisions [8].


In real networks, the modularity value is usually in the range [0.3, 0.7]; 1 means a very strong

group structure and 0 means no better than random.

3. MOBILITY MODEL SETUP

3.1. Mobility Scenario Description

Usually peer nodes in one group act in concert, for example, these nodes may go from office to

restaurant together for lunch, maybe after lunch they will have a rest and then go from restaurant

to cinema for entertainment.

Figure 2 Illustration of mobility scenario

In terms of this realistic mobility scenario, this paper here models the office, restaurant and

cinema to be small squares as shown in figure 2 in simulation scenario.

When the group moves from office (square i ) to restaurant(square j ), the restaurant (square j )

will be regarded as group-destination for all the peer nodes.

When the group continues to move from restaurant the cinema, one new square (square c) in the

simulation will be chosen as the new group-destination corresponding to the cinema.

It should be noted that peer nodes in the same group also have their own self-destination. This

means that these nodes self-destinations may be different positions corresponding to different

points in the group-destination square.

In realistic scenario, another important phenomenon is as follows: when nodes move towards the

destinations, some of them can move fast and some of them perhaps move slowly. This results

that these faster nodes can reach their self-destinations earlier than the slower nodes. After

arriving, these faster nodes would wait other slower nodes for a while before departure for next

new group destination.


3.2. Mobility Model Abstraction

Before moving, each group will randomly choose one small square in the simulation area to be its

group-destination. And then each peer node in the group will choose one position in that small

square to be its self-destination individually. Once they know their self-destinations, each node

will move straightly towards its goal at some speed value. Because of the speed and destination

position difference of nodes in one group, they can’t ensure that all of them will reach their self-

destinations at the same time. So the former arriver will pause for a period of time to wait until

all his peer partners arrive. Upon all the members of the group achieve arrival, they will choose a

new small square to be their next common destination and move towards it.

3.3. Mobility Mathematical Model

Assume there are total N nodes in one group, the th

n node is denoted as node (1 )n n N≤ ≤ . The

beginning square is denoted as j

S , so the beginning position of node n can be denoted as

( 0)n i

P t t S= ∈ . The group-destination square is denoted as j

S , the self-destination of node n in

jS is denoted as ( 0)

n iP t t S= ∈ .

At time ( 0)t t ≥ the position of node n is a function of coordinate ( )dn

X t and ( )dn

Y t which can

be expressed as follows:

cos

sin

( 1) ( ) ( ) ( )( 1)

( 1) ( ) ( ) ( )

n n n n

n n n n

n

X t X t V t tP t

Y t Y t V t t

⋅ Φ

⋅ Φ

+ ++ = =

+ +

Where

t is the current moment;

1t + is the next moment;

( )n

tΦ is the direction function of node n which can be defined as

( )( ) ( )

( )

dn n

n

dn n

Y Y tt arctg

X X tΦ

−=

−

�� is the velocity function which obeys normal distribution.

3.4. Speed Choice

The research in [13] has shown that the walking speed of a pedestrian obeys normal distribution.

The measurement in [14] shows that the mean speed of a walking pedestrian is a range which is

from 1.16 m/s to 1.58 m/s representing walking normally or walking fast.

In [15], the Manual of Uniform Traffic Control Devices (MUTCD) shows that pedestrians walk

with a normal speed of 1.2 m/s (4 ft/sec). [10] indicates a statistics that walking speed for

younger pedestrians (ages 14 to 64) was 1.25 m/sec (4.09 ft/sec); for older pedestrians (ages 65

and over) it was 0.97 m/sec (3.19 ft/sec). For designing purposes values of 1.22 m/sec (4 ft/sec)

for younger pedestrians and 0.91 m/sec (3 ft/sec) for older pedestrians are appropriate[10].


In this paper, the mean value of walking speed of the node n , vnµ , equals to 1.22 m/s [10] and

its standard deviation vnσ is uniformly chosen in the range of [ ]0 0.26 /m s [13]. So the speed

of node n at time t is

( )n

V t ~ ( ),v vn nN µ σ

The probability density function of ( )n

V t can be expressed as:

2

2

( )

( )exp( )

2

2

v

v

v

n

n

n

f v

v µ

σ

σ π=

−−

4. MODEL CHARACTER ANALYSIS

4.1. Classical Individual Mobility Model-RWP

0 0( , )x y

1 1( , )x y

2 2( , )x y

3 3( , )x y

4 4( , )x y

5 5( , )x y6 6( , )x y

7 7( , )x y

8 8( , )x y

Figure 3 Illustration of RWP mobility

Josh et al. [8] present a Random Waypoint model. In this model a mobile node is initially placed

at random location in the simulation area with a destination and speed according to the random

distribution. Then the mobile node moves from its current location to it’s destinations via a

straight line. After arrival at the destination, the mobile node determines a pause time according

to a random distribution, and after that chooses a new destination.

4.2. Analysis and Comparison

The proposed group model in this paper can be regarded as one improvement version of RWP.

To analyze the character of these two mobility models, two important parameters inter-contact

time and contact time are adopted. These two parameters describe the characteristics of

connection opportunities of the network, i.e. how many and when do they occur, how often and

how long. Contact time is defined as the time interval during which the two nodes can keep

contact. The time interval from this contact to next one is defined as inter-contact time during

which nodes can’t communicate. [1] These two parameters are very important for opportunistic


network. Contact time can help us determining the capacity of opportunistic networks. Inter-

contact time is a parameter which strongly affects the feasibility of the opportunistic networking.

The new mobility model in this paper is named as Peer Node Group Mobility Model (PNGM)

here. We let each mobility model run 1000 seconds for one experiment and made 50 times

similar experiments for each mobility model. Figure 4 and Figure 5 give the contact time

distribution and inter-contact time distribution in different coordinate systems. In Fig. 4 (a), we

can see that the new mobility model’ inter-contact time distribution behave an exponential

distribution plot using log-log coordinate system. In Fig. 4 (b), the contact time distributions

show more evident difference. The RWP mobility models’ curves are still exponential-like

curves. In contrast, the shape of PNGM model’ curves are changed. The PNGM model’s curve is

one kind of transition from exponential-like curve to power- law-like curve. This shows PNGM

model behaves stronger group character than RWP model. In Fig. 5 we can see the difference

between the curves more clearly using semi-log coordinate system. The exponential nature of the

inter-contact time shown in Fig. 4 (a) becomes straight line form in Fig. 5 (a).

Fig 4 (a)

Fig4 (b)

Figure 4. Inter-contact time cdf and contact time cdf in log-log coordinates.

100

101

102

103

10-3

10-2

10-1

100

time (second)

P[X

>x]


RWP

PNGM

100

101

102

103

10-3

10-2

10-1

100

time (second)

P[X

>x]

Contact time cdf

RWP

PNGM


Fig 5 (a)

Fig5 (b)

Figure 5. Inter-contact time cdf and contact time cdf in semi-log coordinates.

5. CONCLUSIONS

In this paper, we proposed one new group mobility model with group character. We analyze its

curves about contact time, inter-contact time. The curves show different shapes with the classical

individual model RWP. By comparison, PNGM model shows obvious group character than RWP

model.

ACKNOWLEDGEMENTS

This author would like to thank DFH satellite Co., Ltd. for the support to this paper.

0 200 400 600 800 100010

-3

10-2

10-1

100

time (second)

P[X

>x]


RWP

PNGM

0 200 400 600 800 100010

-3

10-2

10-1

100

time (second)

P[X

>x]

Contact time cdf

RWP

PNGM


REFERENCES

[1] A. Chaintreau, P. Hui, J. Crowcroft, C. Diot, R. Gass, and J. Scott, “Pocket Switched Networks: Real-

world mobility and its consequences for opportunistic forwarding,” Technical Report UCAM-CL-

TR-617, University of Cambridge, Computer Laboratory, February 2005.

[2] J Broch, DA Maltz, DB Johnson, YC Hu, J Jetcheva,” Multi-hop wireless ad hoc network routing

protocols”, in Proceedings of the ACM/IEEE International Conference on Mobile Computing and

Networking (Mobicom 1998), pages 85-97, 1998.

[3] Y. Ko and N.H. Vaidya, “Location-aided routing (LAR) in mobile ad hoc networks,” in Proceedings

of the ACM/IEEE International Conference on Mobile Computing and Networking (Mobicom),

pages 66-75, 1998.

[4] M. E. J. Newman, Detecting community structure in networks. Eur. Phys. J. B, 38, 321–330 (2004)..

[5] E. de Silva and M. Stumpf, “Complex networks and simple models in Biology”, J. R. Soc. Interface,

2 (2005), pp. 419-430..

[6] N. Przulj, D. G. Corneil, and I. Jurisica, “Modeling interactome:Scale- free or geometric?”,

Bioinformatics, 20 (2004), pp. 3508-3515..

[7] D. J. Watts and S. H. Strogatz, “Collective dynamics of 'small-world' networks”, Nature, 393 (1998),

pp. 440-442.

[8] M.E.J. Newman and M. Girvan. “Finding and evaluaing community structure in networks”. Physical

Review E,68,2003.

[9] Mirco Musolesi, Cecilia Mascolo, “A community based mobility model for ad hoc network research”,

in Proceedings of the 2nd international Workshop on Multi-Hop Ad Hoc Networks: From theory To

Reality (REALMAN '06), Florence, Italy, May 2006, pp 31-38.

[10] R.L. Knoblauch, M.T. Pietrucha, M. Nitzburg. Field studies of pedestrian walking speed and start-up

time. Transportation Research Board Records No. 1538, 1996.

[11] Newman, M. E. J. Fast algorithm for detecting community structure in networks. Phys. Rev. E

69,066133 (2004)

[12] Girvan, M. and Newman, M. E. J., Community structure in social and biological networks, Proc.

Natl. Acad.Sci. USA 99, 8271–8276 (2002).

[13] L. Henderson. The statistics of crowd fluids. Nature, Vol. no. 229, 381-383, 1971.

[14] Finnis, K.K. and Walton, D. Field observations of factors influencing walking speeds. Ergonomics,

2006.

[15] LaPlante, John and Kaeser, Thomas P. “A History of Pedestrian Signal Walking Speed

Assumptions”, in Proceedings of the 3rd Urban Street Symposium: Uptown, Downtown, or Small

Town: Designing Urban Streets That Work, 2007.

AUTHORS

GuoDong KANG is an telecommunication engineer of DFH Satellite Co., Ltd., 100094,

Beijing, China.

Guoliang KANG is an PhD student of University of Technology, Sydney 15 Broadway, Ultimo NSW 2007

AUTHOR INDEX

Bohua Gan 37

Chaman L. Sabharwal 27

Feiyang Liu 01

Felming Feng 37

Girija Chetty 73

Guan Wang 37

Guanying Wang 37

GuoDong Kang 13, 81

GuoLiang Kang 13, 81

Haibo Zhang 01

Jennifer L. Leopold 27

Jinzhi Ning 13

Juan Zhang 01

Khairul Azmi Abu Bakar 61

Luming Wan 01

Mohammad Alwadi 73

Mohd Shafeq Md Hasan 61

Naihai Zou 37

Nor Izyani Daud 61

Pil Seong Park 49

Vincent Chang 37

Xiuli Pan 37

computer science & information technology 41 · computer science & information technology...

Documents